Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: start audio.cpp options at each model's recommended values (#13764)

Etherl345217f18595
diff --git a/studio/backend/core/inference/audio_cpp_backend.py b/studio/backend/core/inference/audio_cpp_backend.py
index 56d46218163..4d952693987 100644
--- a/studio/backend/core/inference/audio_cpp_backend.py
+++ b/studio/backend/core/inference/audio_cpp_backend.py
@@ -37,9 +37,10 @@ from core.inference.audio_cpp_models import (
CloneSpec,
forget,
option_matches,
+ filled_options,
require_runnable,
resolve,
- validate_options,
+ workflow_options,
)
from core.inference.audio_cpp_server import (
AudioCppRequestCancelledError,
@@ -143,11 +144,39 @@ _SPEAKER_LINE_RE = re.compile(r"^\s*Speaker\s*\d+\s*:", re.IGNORECASE | re.MULTI
_MIN_SPEED, _MAX_SPEED = 0.5, 2.0
+def _primary_workflow(model: AudioCppModel) -> str:
+ workflows = list(model.workflows)
+ return "speak" if "speak" in workflows or not workflows else workflows[0]
+
+
+def options_by_workflow(model: AudioCppModel, primary: str) -> Optional[dict[str, list[dict]]]:
+ """Each workflow's options where they differ from ``audio_options``; convert always listed."""
+ base = workflow_options(model, primary)
+ out = {
+ name: [dict(option) for option in options]
+ for name in model.workflows
+ if (options := workflow_options(model, name)) != base
+ or (name == "convert" and model.convert is not None)
+ }
+ return out or None
+
+
+def _reread_after_download(model: AudioCppModel, hf_token: Optional[str]) -> AudioCppModel:
+ """``model`` re-resolved from the served file: quants can embed different specs (MagpieTTS
+ ``orig`` says ``speaker``, its Q8_0 ``voice_id``)."""
+ forget(model.id)
+ fresh = resolve(model.id, model.variant.key, hf_token, network = False)
+ if fresh is None or fresh.unsupported or fresh.variant.key != model.variant.key:
+ return model
+ return fresh
+
+
def model_info_fields(model: AudioCppModel) -> dict[str, Any]:
"""What status reports for a loaded audio.cpp model, beyond the common audio fields."""
+ primary = _primary_workflow(model)
return {
"audio_family": model.family,
- "audio_options": [dict(option) for option in model.options],
+ "audio_options": [dict(option) for option in workflow_options(model, primary)],
"gguf_variant": model.variant.key,
"audio_workflows": list(model.workflows),
"audio_reference_text": model.clone.reference_text if model.clone else None,
@@ -155,11 +184,7 @@ def model_info_fields(model: AudioCppModel) -> dict[str, Any]:
"audio_clone": clone_rules(model),
"audio_edit": edit_rules(model),
"audio_music": music_rules(model),
- "audio_options_by_workflow": (
- {"convert": [dict(option) for option in model.convert_options]}
- if model.convert is not None
- else None
- ),
+ "audio_options_by_workflow": options_by_workflow(model, primary),
"audio_workflow_tasks": audio_cpp_convert.workflow_tasks(model),
"audio_convert": audio_cpp_convert.convert_caps(model),
"audio_convert_rules": model.convert
@@ -202,8 +227,7 @@ def edit_rules(model: AudioCppModel) -> Optional[dict[str, Any]]:
def edit_options(model: AudioCppModel) -> tuple[dict, ...]:
- claims = set(model.edit.claims) if model.edit is not None else set()
- return tuple(option for option in model.options if option["name"] not in claims)
+ return workflow_options(model, "edit")
def edit_request_bodies(
@@ -354,7 +378,8 @@ class AudioCppBackend:
raise RuntimeError(problem)
self.loading_models.add(model_name)
try:
- self._ensure_downloaded(model, hf_token)
+ if self._ensure_downloaded(model, hf_token):
+ model = _reread_after_download(model, hf_token)
self._session_overrides = {}
self._status_patch = None
self._start_server(model)
@@ -376,12 +401,14 @@ class AudioCppBackend:
finally:
self.loading_models.discard(model_name)
- def _ensure_downloaded(self, model: AudioCppModel, hf_token: Optional[str]) -> None:
- self._download_missing(model, hf_token)
+ def _ensure_downloaded(self, model: AudioCppModel, hf_token: Optional[str]) -> bool:
+ """Fetch what the model and its companions still lack; True when the model's own files were."""
+ fetched = self._download_missing(model, hf_token)
for companion in model.companions:
companion_model = _resolve_companion(model, companion, hf_token, network = True)
if self._download_missing(companion_model, hf_token):
forget(companion.id)
+ return fetched
@staticmethod
def _download_missing(
@@ -494,7 +521,7 @@ class AudioCppBackend:
if model.task == "sep":
raise RuntimeError(f"{model.display_name} separates audio; open Separate.")
if workflow == "music" and music is not None:
- options = validate_options(model.options, audio_options)
+ options = filled_options(model, "music", audio_options)
return self._run_music(
model,
music,
@@ -517,7 +544,9 @@ class AudioCppBackend:
mode = str(convert.get("mode") or "speech")
try:
served, options = audio_cpp_convert.served_model(
- model, mode, validate_options(model.convert_options, audio_options)
+ model,
+ mode,
+ filled_options(model, "convert", audio_options),
)
request = audio_cpp_convert.convert_request(
model,
@@ -535,15 +564,13 @@ class AudioCppBackend:
except audio_cpp_convert.ConvertRequestError as exc:
raise RuntimeError(str(exc)) from exc
elif editing:
- options = validate_options(edit_options(model), audio_options)
+ options = filled_options(model, "edit", audio_options)
served = _served_for_edit(model)
else:
if cloning and model.clone is None:
raise RuntimeError(f"{model.display_name} cannot clone a voice.")
served = model
- options = validate_options(
- model.clone_options if cloning else model.options, audio_options
- )
+ options = filled_options(model, "clone" if cloning else "speak", audio_options)
server = self._running_server(served, cancel_event)
try:
if converting:
diff --git a/studio/backend/core/inference/audio_cpp_models.py b/studio/backend/core/inference/audio_cpp_models.py
index ab531c59b8d..86629981283 100644
--- a/studio/backend/core/inference/audio_cpp_models.py
+++ b/studio/backend/core/inference/audio_cpp_models.py
@@ -34,6 +34,13 @@ from dataclasses import dataclass, field, replace
from pathlib import Path
from typing import Any, BinaryIO, Iterable, Optional, Sequence
+from core.inference.audio_cpp_option_table import (
+ ACE_STEP_BASE_NAME,
+ ENUM_PRESETS,
+ STUDIO_OPTION_LISTS,
+ OptionRule,
+ rules_for,
+)
from loggers import get_logger
logger = get_logger(__name__)
@@ -718,7 +725,13 @@ _FAMILY_LIST: tuple[AudioCppFamily, ...] = (
AudioCppFamily(
"pocket_tts", "tts", request_defaults = {"voice": "alba"}, clone = CloneSpec("unused")
),
- AudioCppFamily("supertonic", "tts", request_defaults = {"voice": "F1"}),
+ # Supertonic's presets (include/engine/models/supertonic/session.h:22 defaults to M1).
+ AudioCppFamily(
+ "supertonic",
+ "tts",
+ request_defaults = {"voice": "M1"},
+ voices = ("M1", "M2", "M3", "M4", "M5", "F1", "F2", "F3", "F4", "F5"),
+ ),
AudioCppFamily("qwen3_tts", "tts"),
*_CLONE_FAMILIES,
*_CONVERT_FAMILIES,
@@ -1227,6 +1240,7 @@ class AudioCppHeader:
spec: Optional[dict] = field(default = None, hash = False, compare = False)
# False when the read stopped before the header's end (a remote prefix): absent keys may lie beyond it.
complete: bool = True
+ source: Optional[str] = field(default = None, compare = False)
@property
def is_audio_cpp(self) -> bool:
@@ -1900,6 +1914,10 @@ def _clean_option(raw: Any) -> Optional[dict]:
if kind == "enum":
values = raw.get("values")
if not isinstance(values, list) or not values:
+ # A shared domain the spec names instead of listing (text_chunk_mode_full).
+ values = list(ENUM_PRESETS.get(str(raw.get("preset") or ""), ()))
+ # One choice is no choice.
+ if len(values) < 2:
return None
option["values"] = [str(v) for v in values]
for bound in ("min", "max"):
@@ -1933,6 +1951,11 @@ def option_schema(
source = spec
if source is None or not ((source.get("options") or {}).get("request")):
source = embedded
+ if not ((source or {}).get("options") or {}).get("request"):
+ # A legacy spec declares none, though the loader reads some.
+ listed = STUDIO_OPTION_LISTS.get(policy.family)
+ if listed:
+ source = {"options": {"request": list(listed)}, "ui": (source or {}).get("ui")}
raw = ((source or {}).get("options") or {}).get("request") or []
driven = _MUSIC_DRIVEN_OPTIONS if policy.task == "music" else frozenset()
if policy.convert is not None:
@@ -1948,9 +1971,10 @@ def option_schema(
names.add(option["name"])
out.append(option)
if policy.task == "tts" and policy.default_server_task != "vdes":
- ui = (source or {}).get("ui") or (embedded or {}).get("ui") or {}
+ ui = (source or {}).get("ui") or (spec or {}).get("ui") or (embedded or {}).get("ui") or {}
voices = (ui.get("builtin_voices") if isinstance(ui, dict) else None) or list(policy.voices)
- if isinstance(voices, list) and voices and all(isinstance(v, str) for v in voices):
+ # One voice is no choice: the request's own default speaks it.
+ if isinstance(voices, list) and len(voices) > 1 and all(isinstance(v, str) for v in voices):
default = policy.request_defaults.get("voice") or ui.get("default_voice")
out.insert(
0,
@@ -2030,6 +2054,10 @@ class AudioCppModel:
request_defaults: dict = field(default_factory = dict, hash = False, compare = False)
model_options: dict = field(default_factory = dict, hash = False, compare = False)
options: tuple[dict, ...] = field(default = (), hash = False, compare = False)
+ # ``options`` before the family's OPTION_RULES, so a variant change re-applies them from scratch.
+ spec_options: tuple[dict, ...] = field(default = (), hash = False, compare = False)
+ # Option names each workflow ("*" = all) leaves out, from the rules.
+ option_hidden: dict = field(default_factory = dict, hash = False, compare = False)
unsupported: Optional[str] = None
local_path: Optional[str] = None
tags: tuple[str, ...] = ()
@@ -2045,6 +2073,10 @@ class AudioCppModel:
request_keys: Optional[frozenset[str]] = field(default = None, hash = False, compare = False)
from_hub: bool = field(default = False, hash = False, compare = False)
+ def __post_init__(self) -> None:
+ if not self.spec_options and self.options:
+ object.__setattr__(self, "spec_options", tuple(self.options))
+
@property
def is_package(self) -> bool:
return not self.variant.primary
@@ -2115,17 +2147,120 @@ class AudioCppModel:
**variant.session_options,
}
changes = _ace_step_variant(self, variant) if self.family == "ace_step" else {}
- return replace(self, variant = variant, model_options = model_options, **changes)
+ model = replace(self, variant = variant, model_options = model_options, **changes)
+ options, hidden = apply_option_rules(
+ model.family, _rule_text(model), model.spec_options, Path(variant.primary).name
+ )
+ return replace(model, options = options, option_hidden = hidden)
-def _ace_step_variant(model: "AudioCppModel", variant: AudioCppVariant) -> dict[str, Any]:
- """Base takes 1-200 steps (32-64 recommended) and the "complete" task; turbo 1-20 and no
- "complete" (ACE-Step-1.5 acestep/constants.py TASK_TYPES_TURBO vs TASK_TYPES_BASE)."""
- base = bool(re.search(r"(^|/)base(/|$)", f"{variant.key} {variant.primary}".lower()))
- steps = {"max": 200, "default": 32} if base else {"max": 20, "default": 8}
- options = tuple(
- {**o, **steps} if o.get("name") == "num_inference_steps" else o for o in model.options
+def _rule_text(model: "AudioCppModel") -> str:
+ """What an OptionRule ``match`` is searched in: repo id, folder, variant key and file names."""
+ return " ".join((model.id, model.folder, model.variant.key, *model.files)).lower()
+
+
+def apply_option_rules(
+ family: str,
+ text: str,
+ options: Sequence[dict],
+ served: str = "",
+) -> tuple[tuple[dict, ...], dict[str, frozenset[str]]]:
+ """``options`` with the family's matching rules applied, and the names each workflow hides."""
+ rules = rules_for(family, text, served)
+ out = [dict(o) for o in options]
+ hidden: dict[str, frozenset[str]] = {}
+ for rule in rules:
+ names = {o["name"] for o in out}
+ for raw in rule.extra:
+ option = _clean_option(raw)
+ if option is not None and option["name"] not in names:
+ names.add(option["name"])
+ out.append(option)
+ for option in out:
+ _apply_rule(rule, option)
+ for workflow, names_hidden in rule.hidden.items():
+ hidden[workflow] = hidden.get(workflow, frozenset()) | frozenset(names_hidden)
+ for option in out:
+ # A default outside the (possibly narrowed) domain is no default.
+ if option.get("default") is not None and not validate_options(
+ (option,), {option["name"]: option["default"]}
+ ):
+ option["default"] = None
+ return tuple(out), hidden
+
+
+def _apply_rule(rule: OptionRule, option: dict) -> None:
+ name = option["name"]
+ if name in rule.bounds:
+ low, high = rule.bounds[name]
+ if low is not None:
+ option["min"] = low
+ if high is not None:
+ option["max"] = high
+ if name in rule.values and option["type"] == "enum":
+ option["values"] = [v for v in rule.values[name] if v in (option.get("values") or [])]
+ if name in rule.defaults:
+ option["default"] = rule.defaults[name]
+ if option["type"] in ("int", "float") and option.get("default") is not None:
+ number = float(option["default"])
+ if option.get("min") is not None:
+ number = max(float(option["min"]), number)
+ if option.get("max") is not None:
+ number = min(float(option["max"]), number)
+ option["default"] = int(round(number)) if option["type"] == "int" else number
+
+
+def filled_options(model: "AudioCppModel", workflow: str, values: Optional[dict]) -> dict:
+ """``values`` over the model's recommended defaults; tool panels send their own options."""
+ return request_options(
+ workflow_options(model, workflow),
+ values,
+ model.request_keys,
+ fill = (o["name"] for o in model.options),
)
+
+
+def workflow_options(model: "AudioCppModel", workflow: str) -> tuple[dict, ...]:
+ """The options ``workflow`` shows and accepts; only Speak offers a built-in voice."""
+ if workflow == "convert":
+ options = model.convert_options
+ elif workflow == "clone":
+ options = model.clone_options
+ elif workflow == "edit":
+ claims = set(model.edit.claims) if model.edit is not None else set()
+ options = tuple(o for o in model.options if o["name"] not in claims)
+ else:
+ options = model.options
+ hidden = set(model.option_hidden.get("*", ())) | set(model.option_hidden.get(workflow, ()))
+ if workflow in ("clone", "edit", "convert"):
+ hidden.add("voice")
+ return tuple(o for o in options if o["name"] not in hidden)
+
+
+def request_options(
+ schema: Sequence[dict],
+ values: Optional[dict],
+ request_keys: Optional[frozenset[str]] = None,
+ fill: Optional[Iterable[str]] = None,
+) -> dict:
+ """Recommended defaults (names in ``fill``, never ``voice``; only ``request_keys`` when strict)
+ under the user's own values."""
+ fill = None if fill is None else frozenset(fill)
+ defaults = {
+ o["name"]: o["default"]
+ for o in schema
+ if o.get("default") is not None
+ and o["name"] != "voice"
+ and (fill is None or o["name"] in fill)
+ and (request_keys is None or o["name"] in request_keys)
+ }
+ return {**validate_options(schema, defaults), **validate_options(schema, values)}
+
+
+def _ace_step_variant(model: "AudioCppModel", variant: AudioCppVariant) -> dict[str, Any]:
+ """Base takes the "complete" task and turbo does not (ACE-Step-1.5 acestep/constants.py
+ TASK_TYPES_TURBO vs TASK_TYPES_BASE)."""
+ base = bool(re.search(ACE_STEP_BASE_NAME, Path(variant.primary).name, re.IGNORECASE))
music = model.music
if music is not None:
modes = []
@@ -2135,7 +2270,7 @@ def _ace_step_variant(model: "AudioCppModel", variant: AudioCppVariant) -> dict[
m = replace(m, actions = (*actions, "continue") if base else actions)
modes.append(m)
music = replace(music, modes = tuple(modes))
- return {"options": options, "music": music}
+ return {"music": music}
class AudioCppModelError(ValueError):
@@ -2211,6 +2346,7 @@ def _probe_header(
header = read_remote_header(ref.repo_id, f.path, hf_token, size = f.size or None)
if header is None:
continue
+ header = replace(header, source = f.path)
if not header.is_audio_cpp:
return header
if header.family:
@@ -2411,6 +2547,12 @@ def _resolve_uncached(
+ ", ".join(v.key for v in variants)
)
chosen = default
+ if header is not None and header.source and chosen.primary and chosen.primary != header.source:
+ # Quantizations of one model can embed different specs (MagpieTTS's orig names an option
+ # ``speaker`` that its Q8_0 calls ``voice_id``): options must come from the file served.
+ own = _probe_header(ref, [RepoFile(chosen.primary)], hf_token, network)
+ if own is not None and own.is_audio_cpp and own.family == family and own.spec:
+ embedded = own.spec
model = AudioCppModel(
id = ref.id,
repo_id = ref.repo_id,
@@ -2425,7 +2567,7 @@ def _resolve_uncached(
needs_espeak = policy.needs_espeak,
request_defaults = dict(policy.request_defaults),
model_options = dict(policy.model_options),
- options = option_schema(policy, spec, embedded),
+ spec_options = option_schema(policy, spec, embedded),
unsupported = unsupported,
local_path = ref.local_path,
tags = tuple(tags),
diff --git a/studio/backend/core/inference/audio_cpp_option_table.py b/studio/backend/core/inference/audio_cpp_option_table.py
new file mode 100644
index 00000000000..ff1a6a1c4f8
--- /dev/null
+++ b/studio/backend/core/inference/audio_cpp_option_table.py
@@ -0,0 +1,817 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Per-model request options and recommended values for the pinned audio.cpp runtime.
+
+The published GGUFs often embed a spec older than the runtime, and some families declare no request
+options at all even though their loaders read them. ``STUDIO_OPTION_LISTS`` names the options such a
+family reads; it is used only when neither the runtime nor the GGUF declares any. ``OPTION_RULES`` then
+adjusts whichever list a model ends up with: the value Unsloth Studio recommends (shown in Advanced
+and sent when the user leaves it alone), tighter bounds, and options a workflow has no use for.
+
+Recommended values come from audio.cpp's own web UI table (``webui/configs/model_params.json``),
+else the runtime's default in the C++ loader, else the upstream model card; each one names its source.
+A default of ``None`` sends nothing, for options the runtime works out per request. The table holds
+only for the tags it was checked against.
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass, field
+from types import MappingProxyType
+from typing import Any, Mapping, Optional
+
+# The unslothai/audio.cpp releases this table was checked against (v0.9.1-unsloth.1 is the same
+# commit as v0.9.0-unsloth.1).
+AUDIO_CPP_OPTION_TABLE_TAGS = frozenset({"v0.9.0-unsloth.1", "v0.9.1-unsloth.1"})
+
+# Enum domains a spec may name instead of listing (audio.cpp src/framework/model_spec/options.cpp).
+ENUM_PRESETS: Mapping[str, tuple[str, ...]] = MappingProxyType(
+ {
+ "weight_type_full": ("native", "f32", "f16", "bf16", "q8_0"),
+ "weight_type_conv": ("native", "f32", "f16"),
+ "weight_type_codec_q8": ("native", "f32", "f16", "q8_0"),
+ "text_chunk_mode_full": ("word_budget", "tag_aware", "japanese", "endline"),
+ "perf_mode_flash_attention": ("off", "flash_attention"),
+ "best_of_n_language": ("auto", "en", "ja"),
+ }
+)
+
+_EMPTY: Mapping[str, Any] = MappingProxyType({})
+
+
+@dataclass(frozen = True)
+class OptionRule:
+ """How Unsloth Studio adjusts one family's options, for every model or only matching ones."""
+
+ # Regex over repo id, folder, variant key and file names; None = whole family. All matches apply.
+ match: Optional[str] = None
+ # Regex over the served GGUF's file name only.
+ served: Optional[str] = None
+ # Options the family reads but its list lacks, in the spec's request-option shape.
+ extra: tuple[dict, ...] = ()
+ # Recommended value per option name; None sends nothing and leaves the runtime to decide.
+ defaults: Mapping[str, Any] = field(default_factory = lambda: _EMPTY)
+ # (min, max) per option name; None keeps that side as declared.
+ bounds: Mapping[str, tuple[Optional[float], Optional[float]]] = field(
+ default_factory = lambda: _EMPTY
+ )
+ values: Mapping[str, tuple[str, ...]] = field(default_factory = lambda: _EMPTY)
+ # Option names a workflow never shows or sends, keyed by workflow id or "*" for every workflow.
+ hidden: Mapping[str, frozenset[str]] = field(default_factory = lambda: _EMPTY)
+ # Where each recommended value comes from: a model_params.json key, a C++ file:line or a URL.
+ sources: Mapping[str, str] = field(default_factory = lambda: _EMPTY)
+
+
+# audio.cpp loads an ACE-Step GGUF as base/SFT when its file name contains "base" or "sft", else as
+# turbo (src/models/ace_step/loader.cpp selection_from_request; turbo is the default).
+ACE_STEP_BASE_NAME = r"base|sft"
+
+
+def _opt(name: str, type_: str, description: str, **kw: Any) -> dict:
+ return {"name": name, "type": type_, "description": description, **kw}
+
+
+_CHUNK_MODES = ["default", "tag_aware", "japanese", "endline"]
+
+# Shared by the four Qwen3-TTS packages (src/models/qwen3_tts/session.cpp:28-90).
+_QWEN3_TTS_OPTIONS = (
+ _opt("temperature", "float", "Talker sampling temperature.", min = 0.0, max = 2.0),
+ _opt("top_k", "int", "Talker top-k sampling limit.", min = 0),
+ _opt("top_p", "float", "Talker nucleus sampling.", min = 0.0, max = 1.0),
+ _opt("repetition_penalty", "float", "Talker repetition penalty.", min = 1.0, max = 2.0),
+ _opt("do_sample", "bool", "Sample the talker; off decodes greedily."),
+ _opt("subtalker_do_sample", "bool", "Sample the code predictor; off decodes greedily."),
+ _opt(
+ "subtalker_temperature", "float", "Code predictor sampling temperature.", min = 0.0, max = 2.0
+ ),
+ _opt("subtalker_top_k", "int", "Code predictor top-k sampling limit.", min = 0),
+ _opt("subtalker_top_p", "float", "Code predictor nucleus sampling.", min = 0.0, max = 1.0),
+ _opt("max_tokens", "int", "Most codec frames to generate.", min = 1),
+ _opt("text_chunk_size", "int", "Most characters per generated text chunk.", min = 1),
+)
+
+_MOSS_SAMPLING = (
+ _opt("do_sample", "bool", "Sample audio tokens; off decodes greedily."),
+ _opt("temperature", "float", "Audio token sampling temperature.", min = 0.01, max = 2.5),
+ _opt("top_p", "float", "Audio token nucleus sampling.", min = 0.0, max = 1.0),
+ _opt("top_k", "int", "Audio token top-k sampling limit.", min = 0),
+ _opt("repetition_penalty", "float", "Audio token repetition penalty.", min = 1.0, max = 2.0),
+ _opt("text_temperature", "float", "Text token sampling temperature.", min = 0.01),
+ _opt("max_tokens", "int", "Most audio frames to generate.", min = 1),
+)
+
+# Request options a family's loader reads when no spec it ships declares any, keyed by family.
+# Names are the C++ reader's; bounds come from the web UI table or the reader's own checks.
+STUDIO_OPTION_LISTS: Mapping[str, tuple[dict, ...]] = MappingProxyType(
+ {
+ # src/models/pocket_tts/session.cpp:136-176, validated at :243-248.
+ "pocket_tts": (
+ _opt("temperature", "float", "Flow sampling temperature.", min = 0.01),
+ _opt("eos_threshold", "float", "End-of-speech logit threshold."),
+ _opt("frames_after_eos", "int", "Frames kept after end of speech (-1 = auto).", min = -1),
+ _opt("noise_clamp", "float", "Clamp on sampling noise (-1 = off)."),
+ ),
+ # src/models/moss/moss_tts_nano/session.cpp:66-115.
+ "moss_tts_nano": _MOSS_SAMPLING,
+ # src/models/moss/moss_tts_local/session.cpp:193-225.
+ "moss_tts_local": _MOSS_SAMPLING,
+ # src/models/supertonic/session.cpp:190-250; the voice picker comes from the family's voices.
+ "supertonic": (
+ _opt("speed", "float", "Speaking rate.", min = 0.5, max = 2.0),
+ _opt("num_inference_steps", "int", "Flow matching steps.", min = 1),
+ _opt("text_chunk_size", "int", "Most characters per generated text chunk.", min = 1),
+ ),
+ # src/community_models/chatterbox_turbo/session.cpp:19-40, :97.
+ "chatterbox_turbo": (
+ _opt("temperature", "float", "Speech token sampling temperature.", min = 0.0),
+ _opt("top_p", "float", "Speech token nucleus sampling.", min = 0.0, max = 1.0),
+ _opt("top_k", "int", "Speech token top-k sampling limit.", min = 0),
+ _opt("repetition_penalty", "float", "Speech token repetition penalty.", min = 0.0),
+ _opt("text_chunk_size", "int", "Most characters per generated text chunk.", min = 1),
+ ),
+ # src/models/voxcpm2/session.cpp:336-341, :593-665.
+ "voxcpm2": (
+ _opt("num_inference_steps", "int", "Flow matching steps.", min = 1),
+ _opt("guidance_scale", "float", "Classifier-free guidance.", min = 0.0, max = 5.0),
+ _opt("min_tokens", "int", "Fewest audio patches to generate.", min = 0),
+ _opt("max_tokens", "int", "Most audio patches to generate (0 = model limit).", min = 0),
+ _opt("retry_badcase", "bool", "Retry a take that runs far too long."),
+ _opt("retry_badcase_max_times", "int", "Most retries of a bad take.", min = 1),
+ _opt(
+ "retry_badcase_ratio_threshold",
+ "float",
+ "Audio-to-text length ratio that counts as a bad take.",
+ # session.cpp:670-676 rejects a non-positive count or ratio.
+ min = 0.1,
+ ),
+ _opt(
+ "text_chunk_mode", "enum", "How the text is split into chunks.", values = _CHUNK_MODES
+ ),
+ ),
+ "qwen3_tts": _QWEN3_TTS_OPTIONS,
+ # src/models/chatterbox/session.cpp:42-70. Exaggeration and guidance are clone tool options.
+ "chatterbox": (
+ _opt("temperature", "float", "Speech token sampling temperature.", min = 0.0, max = 2.0),
+ _opt(
+ "repetition_penalty", "float", "Speech token repetition penalty.", min = 1.0, max = 2.0
+ ),
+ _opt("min_p", "float", "Speech token min-p sampling.", min = 0.0, max = 1.0),
+ _opt("top_p", "float", "Speech token nucleus sampling.", min = 0.0, max = 1.0),
+ ),
+ # src/models/index_tts2/request.cpp:94-195. The emotion options are clone tool options.
+ "index_tts2": (
+ _opt("use_random_emotion", "bool", "Pick a random emotion; lowers voice fidelity."),
+ _opt("duration_factor", "float", "Length multiplier (>1 slower).", min = 0.5, max = 2.0),
+ _opt("interval_silence_ms", "int", "Silence between text segments (ms).", min = 0),
+ _opt("num_beams", "int", "Beam search width.", min = 1),
+ _opt("max_tokens", "int", "Most mel tokens to generate.", min = 1),
+ _opt("temperature", "float", "Sampling temperature.", min = 0.01),
+ _opt("top_p", "float", "Nucleus sampling.", min = 0.01, max = 1.0),
+ _opt("top_k", "int", "Top-k sampling limit.", min = 1),
+ _opt("repetition_penalty", "float", "Repetition penalty.", min = 0.0),
+ _opt("length_penalty", "float", "Beam search length penalty."),
+ _opt("do_sample", "bool", "Sample tokens; off decodes greedily."),
+ ),
+ # src/models/vevo2/session.cpp:814-823, :1037-1085.
+ "vevo2": (
+ _opt("temperature", "float", "AR sampling temperature.", min = 0.01, max = 2.0),
+ _opt("top_k", "int", "AR top-k sampling limit.", min = 0),
+ _opt("top_p", "float", "AR nucleus sampling.", min = 0.0, max = 1.0),
+ _opt("repetition_penalty", "float", "AR repetition penalty.", min = 0.01),
+ _opt("max_tokens", "int", "Most AR tokens to generate.", min = 1),
+ _opt("num_inference_steps", "int", "Flow matching steps.", min = 1, max = 100),
+ _opt(
+ "audio_chunk_duration_sec",
+ "float",
+ "Split long source audio into chunks this long (0 = off).",
+ min = 0.0,
+ ),
+ _opt("cross_fade_duration_sec", "float", "Cross-fade between audio chunks.", min = 0.0),
+ ),
+ # src/models/higgs_audio_tts/session.cpp:98-130, :249.
+ "higgs_audio_tts": (
+ _opt("temperature", "float", "Sampling temperature.", min = 0.0),
+ _opt("top_p", "float", "Nucleus sampling.", min = 0.0, max = 1.0),
+ _opt("top_k", "int", "Top-k sampling limit.", min = 0),
+ _opt("max_tokens", "int", "Most audio tokens to generate.", min = 1),
+ _opt("text_chunk_size", "int", "Most characters per generated text chunk.", min = 1),
+ ),
+ # src/community_models/moss_voicegen/session.cpp:278-289.
+ "moss_voicegen": (
+ _opt("temperature", "float", "Audio token sampling temperature.", min = 0.0),
+ _opt("top_p", "float", "Audio token nucleus sampling.", min = 0.0, max = 1.0),
+ _opt("top_k", "int", "Audio token top-k sampling limit.", min = 0),
+ _opt("repetition_penalty", "float", "Audio token repetition penalty.", min = 0.0),
+ ),
+ # src/models/omnivoice/session.cpp:77-121.
+ "omnivoice": (
+ _opt("num_inference_steps", "int", "Unmasking steps.", min = 1),
+ _opt("guidance_scale", "float", "Classifier-free guidance.", min = 0.0, max = 5.0),
+ _opt("speed", "float", "Speaking rate.", min = 0.5, max = 2.0),
+ _opt("duration", "float", "Output length in seconds; overrides speed.", min = 0.0),
+ _opt("t_shift", "float", "Timestep shift.", min = 0.0),
+ _opt("denoise", "bool", "Denoise the reference clip."),
+ _opt("preprocess_prompt", "bool", "Clean up the reference clip before cloning."),
+ _opt("postprocess_output", "bool", "Trim silence from the output."),
+ _opt("layer_penalty_factor", "float", "Penalty on deeper codebook layers.", min = 0.0),
+ _opt("position_temperature", "float", "Unmasking position temperature.", min = 0.0),
+ _opt("class_temperature", "float", "Token sampling temperature (0 = greedy).", min = 0.0),
+ _opt(
+ "audio_chunk_duration", "float", "Chunk length for long outputs (seconds).", min = 0.0
+ ),
+ _opt(
+ "audio_chunk_threshold",
+ "float",
+ "Output length that switches chunking on (seconds).",
+ min = 0.0,
+ ),
+ ),
+ # src/models/vibevoice/session.cpp:70-120.
+ "vibevoice": (
+ _opt("num_inference_steps", "int", "Diffusion steps.", min = 1),
+ _opt("guidance_scale", "float", "Classifier-free guidance.", min = 0.0, max = 5.0),
+ _opt("max_length_times", "float", "Output length cap relative to the text.", min = 0.1),
+ _opt("do_sample", "bool", "Sample tokens; off decodes greedily."),
+ _opt("temperature", "float", "Sampling temperature.", min = 0.05, max = 2.0),
+ _opt("top_p", "float", "Nucleus sampling.", min = 0.05, max = 1.0),
+ _opt("top_k", "int", "Top-k sampling limit.", min = 0),
+ ),
+ # The pinned runtime's model_specs/fish_audio.json; the published GGUF embeds none.
+ "fish_audio": (
+ _opt("max_tokens", "int", "Most semantic tokens to generate per chunk.", min = 1),
+ _opt("text_chunk_size", "int", "Most characters per generated text chunk.", min = 1),
+ _opt(
+ "text_chunk_mode",
+ "enum",
+ "How the text is split into chunks.",
+ preset = "text_chunk_mode_full",
+ ),
+ # session.cpp:136 takes (0, 1] only.
+ _opt("top_p", "float", "Nucleus sampling.", min = 0.01, max = 1.0),
+ _opt("top_k", "int", "Top-k sampling limit.", min = 1),
+ # session.cpp:142 takes (0, 2) only.
+ _opt(
+ "temperature", "float", "Semantic token sampling temperature.", min = 0.01, max = 1.99
+ ),
+ ),
+ # src/models/miotts/session.cpp:131-180.
+ "miotts": (
+ _opt("temperature", "float", "Sampling temperature.", min = 0.0, max = 2.0),
+ _opt("top_p", "float", "Nucleus sampling.", min = 0.0, max = 1.0),
+ _opt("top_k", "int", "Top-k sampling limit.", min = 0),
+ _opt("repetition_penalty", "float", "Repetition penalty.", min = 1.0, max = 1.5),
+ _opt("presence_penalty", "float", "Presence penalty.", min = 0.0, max = 1.0),
+ _opt("frequency_penalty", "float", "Frequency penalty.", min = 0.0, max = 1.0),
+ _opt("max_tokens", "int", "Most speech tokens to generate per chunk.", min = 1),
+ _opt("do_sample", "bool", "Sample tokens; off decodes greedily."),
+ ),
+ }
+)
+
+OPTION_RULES: Mapping[str, tuple[OptionRule, ...]] = MappingProxyType(
+ {
+ "ace_step": (
+ OptionRule(
+ extra = (
+ _opt("lm_temperature", "float", "Planner LM sampling temperature.", min = 0.0),
+ _opt("lm_cfg_scale", "float", "Planner LM guidance scale.", min = 0.0),
+ _opt("lm_top_p", "float", "Planner LM nucleus sampling.", min = 0.0, max = 1.0),
+ _opt("lm_top_k", "int", "Planner LM top-k (0 = off).", min = 0),
+ ),
+ defaults = {
+ "lm_temperature": 0.85,
+ "lm_cfg_scale": 2.0,
+ "lm_top_p": 0.9,
+ "lm_top_k": 0,
+ },
+ sources = {
+ "lm_temperature": "audio.cpp include/engine/models/ace_step/types.h:31-42",
+ "lm_cfg_scale": "audio.cpp include/engine/models/ace_step/types.h:31-42",
+ "lm_top_p": "audio.cpp include/engine/models/ace_step/types.h:31-42",
+ "lm_top_k": "audio.cpp include/engine/models/ace_step/types.h:31-42",
+ },
+ ),
+ # Turbo (anything but base): the runtime caps steps at 8 and skips guidance.
+ OptionRule(
+ served = rf"^(?!.*(?:{ACE_STEP_BASE_NAME}))",
+ defaults = {"num_inference_steps": 8, "shift": 3.0},
+ bounds = {"num_inference_steps": (1, 8)},
+ hidden = {"*": frozenset({"guidance_scale"})},
+ sources = {
+ "num_inference_steps": "model_params.json ace_step; audio.cpp src/models/ace_step/session.cpp:79-80",
+ "shift": "model_params.json ace_step; https://github.com/ace-step/ACE-Step-1.5 docs/en/INFERENCE.md",
+ },
+ ),
+ # Base takes 1-200 steps, 32-64 recommended, with guidance on.
+ OptionRule(
+ served = ACE_STEP_BASE_NAME,
+ defaults = {"num_inference_steps": 32, "guidance_scale": 7.0, "shift": 1.0},
+ bounds = {"num_inference_steps": (1, 200)},
+ sources = {
+ "num_inference_steps": "https://github.com/ace-step/ACE-Step-1.5 docs/en/INFERENCE.md",
+ "guidance_scale": "https://github.com/ace-step/ACE-Step-1.5 docs/en/INFERENCE.md",
+ "shift": "audio.cpp include/engine/models/ace_step/types.h",
+ },
+ ),
+ ),
+ # Pocket TTS: the runtime picks temperature per package (0.3 english, else 0.7).
+ "pocket_tts": (
+ OptionRule(
+ defaults = {
+ "temperature": None,
+ "eos_threshold": -4.0,
+ "frames_after_eos": -1,
+ "noise_clamp": None,
+ },
+ sources = {
+ "eos_threshold": "audio.cpp include/engine/models/pocket_tts/types.h:55",
+ "frames_after_eos": "model_params.json pocket_tts",
+ },
+ ),
+ ),
+ "moss_tts_nano": (
+ OptionRule(
+ defaults = {
+ "do_sample": True,
+ "temperature": 1.7,
+ "top_p": 0.8,
+ "top_k": 25,
+ "repetition_penalty": 1.0,
+ "text_temperature": 1.5,
+ "max_tokens": None,
+ },
+ sources = {
+ "do_sample": "model_params.json moss_tts_nano",
+ "temperature": "model_params.json moss_tts_nano",
+ "top_p": "model_params.json moss_tts_nano",
+ "top_k": "model_params.json moss_tts_nano",
+ "repetition_penalty": "model_params.json moss_tts_nano",
+ "text_temperature": "audio.cpp include/engine/models/moss/moss_tts_nano/types.h:13",
+ },
+ ),
+ ),
+ # Supertonic: chunk size is 300 characters, 120 for Korean and Japanese.
+ "supertonic": (
+ OptionRule(
+ defaults = {"speed": 1.05, "num_inference_steps": 8, "text_chunk_size": None},
+ sources = {
+ "speed": "model_params.json supertonic (speaking_rate); audio.cpp include/engine/models/supertonic/session.h:20",
+ "num_inference_steps": "model_params.json supertonic",
+ },
+ ),
+ ),
+ "chatterbox_turbo": (
+ OptionRule(
+ defaults = {
+ "temperature": 0.8,
+ "top_p": 0.95,
+ "top_k": 1000,
+ "repetition_penalty": 1.2,
+ "text_chunk_size": 128,
+ },
+ sources = {
+ "temperature": "audio.cpp include/engine/community_models/chatterbox_turbo/tts.h:18",
+ "top_p": "audio.cpp include/engine/community_models/chatterbox_turbo/tts.h:19",
+ "top_k": "audio.cpp include/engine/community_models/chatterbox_turbo/tts.h:20",
+ "repetition_penalty": "audio.cpp include/engine/community_models/chatterbox_turbo/tts.h:21",
+ "text_chunk_size": "audio.cpp src/community_models/chatterbox_turbo/session.cpp:17",
+ },
+ ),
+ ),
+ "voxcpm2": (
+ OptionRule(
+ defaults = {
+ "num_inference_steps": 10,
+ "guidance_scale": 2.0,
+ "min_tokens": 2,
+ "max_tokens": None,
+ "retry_badcase": True,
+ "retry_badcase_max_times": 3,
+ "retry_badcase_ratio_threshold": 6.0,
+ "text_chunk_mode": "tag_aware",
+ },
+ sources = {
+ "num_inference_steps": "model_params.json voxcpm2",
+ "guidance_scale": "model_params.json voxcpm2",
+ "min_tokens": "model_params.json voxcpm2",
+ "retry_badcase": "model_params.json voxcpm2",
+ "retry_badcase_max_times": "audio.cpp include/engine/models/voxcpm2/types.h:22",
+ "retry_badcase_ratio_threshold": "audio.cpp include/engine/models/voxcpm2/types.h:23",
+ "text_chunk_mode": "model_params.json voxcpm2",
+ },
+ ),
+ ),
+ # Qwen3-TTS: max_tokens comes from the package's generation_config.json.
+ "qwen3_tts": (
+ OptionRule(
+ defaults = {
+ "temperature": 0.9,
+ "top_k": 50,
+ "top_p": 1.0,
+ "repetition_penalty": 1.05,
+ "do_sample": True,
+ "subtalker_do_sample": True,
+ "subtalker_temperature": 0.9,
+ "subtalker_top_k": 50,
+ "subtalker_top_p": 1.0,
+ "max_tokens": None,
+ "text_chunk_size": 8192,
+ },
+ sources = {
+ "temperature": "model_params.json qwen3_tts",
+ "top_k": "model_params.json qwen3_tts",
+ "top_p": "model_params.json qwen3_tts",
+ "repetition_penalty": "model_params.json qwen3_tts",
+ "do_sample": "model_params.json qwen3_tts",
+ "subtalker_do_sample": "audio.cpp include/engine/models/qwen3_tts/types.h:26",
+ "subtalker_temperature": "audio.cpp include/engine/models/qwen3_tts/types.h:31",
+ "subtalker_top_k": "audio.cpp include/engine/models/qwen3_tts/types.h:32",
+ "subtalker_top_p": "audio.cpp include/engine/models/qwen3_tts/types.h:33",
+ "text_chunk_size": "audio.cpp src/models/qwen3_tts/session.cpp:20",
+ },
+ ),
+ ),
+ # Chatterbox converts in a separate vc session that reads none of the clone sampling options.
+ "chatterbox": (
+ OptionRule(
+ defaults = {
+ "temperature": 0.8,
+ "repetition_penalty": 1.2,
+ "min_p": 0.05,
+ "top_p": 1.0,
+ },
+ hidden = {
+ "convert": frozenset({"temperature", "repetition_penalty", "min_p", "top_p"})
+ },
+ sources = {
+ "temperature": "model_params.json chatterbox",
+ "repetition_penalty": "model_params.json chatterbox",
+ "min_p": "audio.cpp include/engine/models/chatterbox/tts.h:24",
+ "top_p": "audio.cpp include/engine/models/chatterbox/tts.h:25",
+ },
+ ),
+ ),
+ "index_tts2": (
+ OptionRule(
+ defaults = {
+ "use_random_emotion": False,
+ "duration_factor": 1.0,
+ "interval_silence_ms": 200,
+ "num_beams": 3,
+ "max_tokens": 1500,
+ "temperature": 0.8,
+ "top_p": 0.8,
+ "top_k": 30,
+ "repetition_penalty": 10.0,
+ "length_penalty": 0.0,
+ "do_sample": True,
+ },
+ sources = {
+ "use_random_emotion": "model_params.json index_tts2",
+ "duration_factor": "model_params.json index_tts2",
+ "interval_silence_ms": "model_params.json index_tts2",
+ "num_beams": "audio.cpp include/engine/models/index_tts2/types.h:143",
+ "max_tokens": "audio.cpp include/engine/models/index_tts2/types.h:145",
+ "temperature": "audio.cpp include/engine/models/index_tts2/types.h:141",
+ "top_p": "audio.cpp include/engine/models/index_tts2/types.h:139",
+ "top_k": "audio.cpp include/engine/models/index_tts2/types.h:140",
+ "repetition_penalty": "audio.cpp include/engine/models/index_tts2/types.h:144",
+ "length_penalty": "audio.cpp include/engine/models/index_tts2/types.h:142",
+ "do_sample": "audio.cpp include/engine/models/index_tts2/types.h:138",
+ },
+ ),
+ ),
+ # CosyVoice3 sizes both token limits from the text (20x and 2x) when they are unset.
+ "cosyvoice3": (OptionRule(defaults = {"max_tokens": None, "min_tokens": None}),),
+ # DotTTS-Edit: these only mean something on Edit, which claims them.
+ "dots_tts": (
+ OptionRule(
+ match = r"(^|[-_ /.])edit([-_ ./]|$)",
+ hidden = {
+ workflow: frozenset(
+ {"template_name", "source_text", "target_text", "use_xvector"}
+ )
+ for workflow in ("speak", "clone")
+ },
+ ),
+ ),
+ # Vevo2 chunks source audio only, which a Clone never has.
+ "vevo2": (
+ OptionRule(
+ defaults = {
+ "temperature": 0.7,
+ "top_k": 20,
+ "top_p": 0.8,
+ "repetition_penalty": 1.1,
+ "max_tokens": 500,
+ "num_inference_steps": 32,
+ "audio_chunk_duration_sec": 0.0,
+ "cross_fade_duration_sec": 1.0,
+ },
+ hidden = {
+ "clone": frozenset({"audio_chunk_duration_sec", "cross_fade_duration_sec"})
+ },
+ sources = {
+ "temperature": "model_params.json vevo2",
+ "top_k": "model_params.json vevo2",
+ "top_p": "model_params.json vevo2",
+ "repetition_penalty": "audio.cpp include/engine/models/vevo2/types.h:36",
+ "max_tokens": "audio.cpp include/engine/models/vevo2/types.h:37",
+ "num_inference_steps": "model_params.json vevo2",
+ "audio_chunk_duration_sec": "model_params.json vevo2",
+ "cross_fade_duration_sec": "model_params.json vevo2",
+ },
+ ),
+ ),
+ # FireRedAudio reads these only for ASR (session.cpp:140-185); Clone has one template.
+ "firered_audio": (
+ OptionRule(
+ hidden = {
+ "*": frozenset(
+ {"max_new_tokens", "enable_thinking", "top_k", "top_p", "temperature"}
+ ),
+ "clone": frozenset({"template_name"}),
+ },
+ ),
+ ),
+ # Base ignores template_name; on Instruct every other template breaks a Clone.
+ "fireredtts3": (OptionRule(hidden = {"*": frozenset({"template_name"})}),),
+ # Offered on Speak only, so there is no reference to trim.
+ "dramabox": (OptionRule(hidden = {"*": frozenset({"reference_duration_sec"})}),),
+ # Irodori: no_ref follows the workflow (a Clone sends a reference); the runtime predicts the
+ # duration, keeps the three guidance scales unless overridden and sizes chunks from the model.
+ "irodori_tts": (
+ OptionRule(
+ defaults = {"duration_sec": None, "guidance_scale": None, "text_chunk_size": None},
+ hidden = {"*": frozenset({"no_ref"})},
+ ),
+ # The 500M v3 base checkpoint is not caption-conditioned.
+ OptionRule(
+ match = r"500m",
+ hidden = {"*": frozenset({"caption", "caption_guidance_scale"})},
+ ),
+ ),
+ "higgs_audio_tts": (
+ OptionRule(
+ defaults = {
+ "temperature": 0.8,
+ "top_p": 0.8,
+ "top_k": 30,
+ "max_tokens": 2048,
+ "text_chunk_size": 1024,
+ },
+ sources = {
+ "temperature": "audio.cpp src/models/higgs_audio_tts/session.cpp:103",
+ "top_p": "audio.cpp src/models/higgs_audio_tts/session.cpp:104",
+ "top_k": "audio.cpp src/models/higgs_audio_tts/session.cpp:105",
+ "max_tokens": "audio.cpp src/models/higgs_audio_tts/session.cpp:102",
+ "text_chunk_size": "audio.cpp src/models/higgs_audio_tts/session.cpp:22",
+ },
+ ),
+ ),
+ "moss_tts_local": (
+ OptionRule(
+ defaults = {
+ "do_sample": True,
+ "temperature": 1.7,
+ "top_p": 0.8,
+ "top_k": 25,
+ "repetition_penalty": 1.0,
+ "max_tokens": 4096,
+ "text_temperature": 1.0,
+ },
+ sources = {
+ "do_sample": "model_params.json moss_tts_local",
+ "temperature": "model_params.json moss_tts_local",
+ "top_p": "model_params.json moss_tts_local",
+ "top_k": "model_params.json moss_tts_local",
+ "repetition_penalty": "model_params.json moss_tts_local",
+ "max_tokens": "audio.cpp include/engine/models/moss/moss_tts_local/generator.h:24",
+ "text_temperature": "audio.cpp include/engine/models/moss/moss_tts_local/generator.h:30",
+ },
+ ),
+ ),
+ "moss_voicegen": (
+ OptionRule(
+ defaults = {"temperature": 1.5, "top_p": 0.6, "top_k": 50, "repetition_penalty": 1.1},
+ sources = {
+ "temperature": "model_params.json moss_voicegen",
+ "top_p": "model_params.json moss_voicegen",
+ "top_k": "model_params.json moss_voicegen",
+ "repetition_penalty": "model_params.json moss_voicegen",
+ },
+ ),
+ ),
+ # OmniVoice: an unset duration follows speed.
+ "omnivoice": (
+ OptionRule(
+ defaults = {
+ "num_inference_steps": 32,
+ "guidance_scale": 2.0,
+ "speed": 1.0,
+ "duration": None,
+ "t_shift": 0.1,
+ "denoise": True,
+ "preprocess_prompt": True,
+ "postprocess_output": True,
+ "layer_penalty_factor": 5.0,
+ "position_temperature": 5.0,
+ "class_temperature": 0.0,
+ "audio_chunk_duration": 15.0,
+ "audio_chunk_threshold": 30.0,
+ },
+ sources = {
+ "num_inference_steps": "model_params.json omnivoice",
+ "guidance_scale": "model_params.json omnivoice",
+ "speed": "model_params.json omnivoice",
+ "t_shift": "audio.cpp include/engine/models/omnivoice/types.h:25",
+ "denoise": "audio.cpp include/engine/models/omnivoice/types.h:26",
+ "preprocess_prompt": "audio.cpp include/engine/models/omnivoice/types.h:31",
+ "postprocess_output": "audio.cpp include/engine/models/omnivoice/types.h:27",
+ "layer_penalty_factor": "audio.cpp include/engine/models/omnivoice/types.h:28",
+ "position_temperature": "audio.cpp include/engine/models/omnivoice/types.h:29",
+ "class_temperature": "audio.cpp include/engine/models/omnivoice/types.h:30",
+ "audio_chunk_duration": "audio.cpp include/engine/models/omnivoice/types.h:32",
+ "audio_chunk_threshold": "audio.cpp include/engine/models/omnivoice/types.h:33",
+ },
+ ),
+ ),
+ # VibeVoice: unset steps fall back to the package config (20); the web UI runs 10.
+ "vibevoice": (
+ OptionRule(
+ defaults = {
+ "num_inference_steps": 10,
+ "guidance_scale": 1.3,
+ "max_length_times": 2.0,
+ "do_sample": False,
+ "temperature": 1.0,
+ "top_p": 1.0,
+ "top_k": 50,
+ },
+ sources = {
+ "num_inference_steps": "model_params.json vibevoice",
+ "guidance_scale": "model_params.json vibevoice",
+ "max_length_times": "model_params.json vibevoice",
+ "do_sample": "model_params.json vibevoice",
+ "temperature": "model_params.json vibevoice",
+ "top_p": "model_params.json vibevoice",
+ "top_k": "audio.cpp include/engine/models/vibevoice/types.h:20",
+ },
+ ),
+ ),
+ # VoxCPM1: the embedded spec's defaults are older than the runtime's; prompt_text is the
+ # Clone transcript under another name.
+ "voxcpm1": (
+ OptionRule(
+ defaults = {
+ "num_inference_steps": 10,
+ "guidance_scale": 2.0,
+ "min_tokens": 2,
+ "max_tokens": 4096,
+ "retry_badcase": True,
+ "retry_badcase_max_times": 3,
+ "retry_badcase_ratio_threshold": 6.0,
+ },
+ hidden = {"*": frozenset({"prompt_text"})},
+ sources = {
+ "num_inference_steps": "model_params.json voxcpm1",
+ "guidance_scale": "model_params.json voxcpm1",
+ "min_tokens": "model_params.json voxcpm1",
+ "max_tokens": "audio.cpp include/engine/community_models/voxcpm1/types.h:14",
+ "retry_badcase": "model_params.json voxcpm1",
+ "retry_badcase_max_times": "audio.cpp include/engine/community_models/voxcpm1/types.h:18",
+ "retry_badcase_ratio_threshold": "audio.cpp include/engine/community_models/voxcpm1/types.h:19",
+ },
+ ),
+ ),
+ "fish_audio": (
+ OptionRule(
+ defaults = {
+ "max_tokens": 1024,
+ "text_chunk_size": 200,
+ "text_chunk_mode": "word_budget",
+ "top_p": 0.8,
+ "top_k": 30,
+ "temperature": 0.8,
+ },
+ sources = {
+ name: "audio.cpp model_specs/fish_audio.json"
+ for name in (
+ "max_tokens",
+ "text_chunk_size",
+ "text_chunk_mode",
+ "top_p",
+ "top_k",
+ "temperature",
+ )
+ },
+ ),
+ ),
+ "miotts": (
+ OptionRule(
+ defaults = {
+ "temperature": 0.8,
+ "top_p": 1.0,
+ "top_k": 50,
+ "repetition_penalty": 1.0,
+ "presence_penalty": 0.0,
+ "frequency_penalty": 0.0,
+ "max_tokens": 700,
+ "do_sample": True,
+ },
+ sources = {
+ "temperature": "model_params.json miotts",
+ "top_p": "model_params.json miotts",
+ "top_k": "model_params.json miotts",
+ "repetition_penalty": "model_params.json miotts",
+ "presence_penalty": "audio.cpp include/engine/models/miotts/assets.h:37",
+ "frequency_penalty": "audio.cpp include/engine/models/miotts/assets.h:38",
+ "max_tokens": "audio.cpp src/models/miotts/assets.cpp:44",
+ "do_sample": "audio.cpp include/engine/models/miotts/assets.h:32",
+ },
+ ),
+ ),
+ # YuE2: guidance depends on cot (1.0 for melody/full, 1.01 for off), so it stays unset.
+ "yue2": (
+ OptionRule(
+ extra = (
+ _opt(
+ "semantic_repetition_penalty",
+ "float",
+ "Semantic codec repetition penalty.",
+ min = 0.1,
+ max = 2.0,
+ ),
+ _opt(
+ "abc_temperature",
+ "float",
+ "ABC planning sampling temperature.",
+ min = 0.0,
+ max = 2.0,
+ ),
+ _opt("abc_top_p", "float", "ABC planning nucleus sampling.", min = 0.01, max = 1.0),
+ _opt("abc_top_k", "int", "ABC planning top-k sampling limit.", min = 1),
+ _opt(
+ "abc_repetition_penalty",
+ "float",
+ "ABC planning repetition penalty.",
+ min = 0.1,
+ max = 2.0,
+ ),
+ ),
+ defaults = {
+ "guidance_scale": None,
+ "semantic_temperature": 1.0,
+ "semantic_top_p": 0.95,
+ "semantic_top_k": 100,
+ "semantic_repetition_penalty": 1.2,
+ "abc_temperature": 0.7,
+ "abc_top_p": 0.9,
+ "abc_top_k": 30,
+ "abc_repetition_penalty": 1.005,
+ },
+ sources = {
+ name: "model_params.json yue2"
+ for name in (
+ "semantic_temperature",
+ "semantic_top_p",
+ "semantic_top_k",
+ "semantic_repetition_penalty",
+ "abc_temperature",
+ "abc_top_p",
+ "abc_top_k",
+ "abc_repetition_penalty",
+ )
+ },
+ ),
+ ),
+ # HeartMuLa: infinite mode and its two knobs mean nothing under the music form's length cap.
+ "heartmula": (
+ OptionRule(
+ hidden = {
+ "*": frozenset(
+ {"infinite_mode", "text_chunk_size", "infinite_chunk_audio_duration_ms"}
+ )
+ },
+ ),
+ ),
+ # ControlFoley: masking applies to a video clip, which the music form never sends.
+ "controlfoley": (OptionRule(hidden = {"*": frozenset({"mask_away_clip"})}),),
+ }
+)
+
+
+def rules_for(
+ family: str,
+ text: str,
+ served: str = "",
+) -> tuple[OptionRule, ...]:
+ """The rules of ``family`` that apply to a model described by ``text``, serving file ``served``."""
+ return tuple(
+ rule
+ for rule in OPTION_RULES.get(family, ())
+ if (rule.match is None or re.search(rule.match, text, re.IGNORECASE))
+ and (rule.served is None or re.search(rule.served, served, re.IGNORECASE))
+ )
diff --git a/studio/backend/tests/fixtures/audio_cpp_catalog_specs.json b/studio/backend/tests/fixtures/audio_cpp_catalog_specs.json
new file mode 100644
index 00000000000..29d8ead6f59
--- /dev/null
+++ b/studio/backend/tests/fixtures/audio_cpp_catalog_specs.json
@@ -0,0 +1,63 @@
+{
+ "audio_cpp_tag": "v0.9.0-unsloth.1",
+ "models": {
+ "audio-cpp/MiniMax-Music3-GGUF": {"family":null,"files":[".gitattributes","LICENSE","README.md","condition_encoder.gguf","config.json","config/condition_encoder.json","config/language_model.json","config/rvq_depth_decoder.json","config/transformer.json","config/vocoder.json","language_model_bf16.gguf","language_model_q4_0.gguf","language_model_q4_k.gguf","language_model_q8_0.gguf","rvq_depth_decoder_bf16.gguf","rvq_depth_decoder_q4_k.gguf","rvq_depth_decoder_q8_0.gguf","tokenizer/tokenizer.json","tokenizer/tokenizer_config.json","transformer_bf16.gguf","transformer_q4_0.gguf","transformer_q4_k.gguf","transformer_q8_0.gguf","vocoder.gguf"],"spec":null,"task":"music"},
+ "audio-cpp/Yue2-3B-GGUF": {"family":"yue2","files":[".gitattributes","README.md","examples/demo1.wav","examples/demo1_q4_0.wav","examples/demo1_q8.wav","examples/demo2.wav","examples/demo2_q4_0.wav","examples/demo2_q8.wav","examples/demo3.wav","examples/demo3_q4_0.wav","examples/demo3_q8.wav","examples/demo4.wav","examples/demo4_q4_0.wav","examples/demo4_q8.wav","examples/tonight-awake-audiocpp-current.wav","examples/tonight-awake-audiocpp-current_q4_0.wav","examples/tonight-awake-audiocpp-current_q8.wav","examples/yue2_demo.mp4","lora/README.md","lora/SHA256SUMS","lora/samples/harbor-keeps-the-dawn-cot-off.wav","lora/samples/harbor-keeps-the-dawn-lyrics.txt","lora/samples/harbor-keeps-the-dawn-style.txt","lora/training-metrics.json","sidecars/yue2-generation-config.json","sidecars/yue2-model-config.json","sidecars/yue2-qwen.tiktoken","sidecars/yue2-vae-config.json","yue2-3b-bf16.gguf","yue2-3b-ios-q4_0.gguf","yue2-3b-q4_0.gguf","yue2-3b-q8_0.gguf","yue2-vae-f16.gguf","yue2-vae-f32.gguf"],"spec":{"options":{"request":[{"name":"style","required":true,"type":"string"},{"name":"lyrics","required":false,"type":"string"},{"default":"full","name":"cot","required":false,"type":"enum","values":["off","melody","full"]},{"default":"audio","name":"stop_after","required":false,"type":"enum","values":["abc","semantic","audio"]},{"name":"abc","required":false,"type":"string"},{"name":"abc_file","required":false,"type":"string"},{"name":"semantic_prefix","required":false,"type":"string"},{"name":"semantic_prefix_file","required":false,"type":"string"},{"name":"nar_noise_file","required":false,"type":"string"},{"default":false,"name":"export_semantic","required":false,"type":"bool"},{"max":20.0,"min":0.0,"name":"guidance_scale","required":false,"type":"float"},{"default":1234,"min":0,"name":"seed","required":false,"type":"int"},{"default":8,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"max":5.0,"min":0.0,"name":"abc_temperature","required":false,"type":"float"},{"max":1.0,"min":0.0,"name":"abc_top_p","required":false,"type":"float"},{"min":1,"name":"abc_top_k","required":false,"type":"int"},{"min":0.001,"name":"abc_repetition_penalty","required":false,"type":"float"},{"min":1,"name":"abc_penalty_window","required":false,"type":"int"},{"min":0,"name":"abc_min_tokens","required":false,"type":"int"},{"min":1,"name":"abc_max_tokens","required":false,"type":"int"},{"max":5.0,"min":0.0,"name":"semantic_temperature","required":false,"type":"float"},{"max":1.0,"min":0.0,"name":"semantic_top_p","required":false,"type":"float"},{"min":1,"name":"semantic_top_k","required":false,"type":"int"},{"min":0.001,"name":"semantic_repetition_penalty","required":false,"type":"float"},{"min":1,"name":"semantic_penalty_window","required":false,"type":"int"},{"min":0,"name":"semantic_min_tokens","required":false,"type":"int"},{"min":1,"name":"semantic_max_tokens","required":false,"type":"int"}]},"schema_version":1,"tasks":["music"]},"task":"music"},
+ "audio-cpp/audio.cpp-gguf/ACE-Step1.5-GGUF": {"family":"ace_step","files":["base/ace-step-1.5-base-bf16.gguf","base/ace-step-1.5-base-q8_0.gguf","turbo/ace-step-1.5-turbo-bf16.gguf","turbo/ace-step-1.5-turbo-q8_0.gguf"],"spec":{},"task":"music"},
+ "audio-cpp/audio.cpp-gguf/BS-RoFormer-ep368-GGUF": {"family":"bs_roformer","files":["bs-roformer-ep368-q8_0.gguf"],"spec":{},"task":"sep"},
+ "audio-cpp/audio.cpp-gguf/Breeze-TTS-2-GGUF": {"family":"breeze_tts","files":["LICENSE","NOTICE","breeze-tts-2-bf16.gguf","breeze-tts-2-q4_0.gguf","breeze-tts-2-q8_0.gguf"],"spec":{"options":{"request":[{"default":"Speak clearly and naturally.","name":"instruction","required":false,"type":"string"},{"name":"reference_text","required":false,"type":"string"},{"default":600,"min":1,"name":"text_chunk_size","required":false,"type":"int"},{"default":"default","name":"text_chunk_mode","required":false,"type":"enum","values":["default","tag_aware","japanese","endline"]},{"default":1500,"min":1,"name":"max_tokens","required":false,"type":"int"},{"default":1,"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":0.9,"min":0,"name":"temperature","required":false,"type":"float"},{"default":0.9,"min":0,"name":"depth_temperature","required":false,"type":"float"},{"default":50,"min":0,"name":"top_k","required":false,"type":"int"},{"default":1,"max":1,"min":0,"name":"top_p","required":false,"type":"float"},{"default":0,"min":0,"name":"seed","required":false,"type":"int"}]},"schema_version":1,"tasks":["tts","clone"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Chatterbox-GGUF": {"family":"chatterbox","files":["chatterbox-f16.gguf","chatterbox-q8_0.gguf"],"spec":{},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Chatterbox-Turbo-GGUF": {"family":"chatterbox_turbo","files":["chatterbox-turbo-q8_0.gguf"],"spec":{"tasks":["tts"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Confucius4-TTS-GGUF": {"family":"confucius4_tts","files":["confucius4-tts-orig.gguf"],"spec":{"options":{"request":[{"default":"zh","name":"language","required":false,"type":"string"},{"default":0.8,"min":0,"name":"temperature","required":false,"type":"float"},{"default":0.8,"max":1,"min":0,"name":"top_p","required":false,"type":"float"},{"default":30,"min":1,"name":"top_k","required":false,"type":"int"},{"default":3,"min":1,"name":"num_beams","required":false,"type":"int"},{"default":10,"min":0,"name":"repetition_penalty","required":false,"type":"float"},{"default":1520,"min":1,"name":"max_tokens","required":false,"type":"int"},{"default":25,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"default":0.7,"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":80,"min":1,"name":"text_chunk_size","required":false,"type":"int"},{"default":"default","name":"text_chunk_mode","required":false,"type":"enum","values":["default","tag_aware","japanese","endline"]},{"default":0.3,"min":0,"name":"cross_fade_duration_sec","required":false,"type":"float"},{"default":0.1,"min":0,"name":"edge_fade_duration_sec","required":false,"type":"float"},{"default":0.1,"min":0,"name":"edge_pad_duration_sec","required":false,"type":"float"},{"default":1234,"min":0,"name":"seed","required":false,"type":"int"}]},"schema_version":1,"tasks":["clone"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/ControlFoley-GGUF": {"family":"controlfoley","files":["controlfoley-large-44k-f32.gguf","controlfoley-large-44k-q8_0.gguf"],"spec":{"options":{"request":[{"default":8,"max":60,"min":0.64,"name":"duration_sec","required":false,"type":"float"},{"default":25,"max":200,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"default":4.5,"max":20,"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":"","name":"negative_prompt","required":false,"type":"string"},{"name":"video","required":false,"type":"path"},{"default":false,"name":"mask_away_clip","required":false,"type":"bool"},{"default":42,"max":2147483647,"min":0,"name":"seed","required":false,"type":"int"}]},"schema_version":1,"tasks":["sfx"]},"task":"music"},
+ "audio-cpp/audio.cpp-gguf/CosyVoice3-GGUF": {"family":"cosyvoice3","files":["cosyvoice3-f32.gguf","cosyvoice3-q8_0.gguf"],"spec":{"options":{"request":[{"default":"zero_shot","name":"template_name","required":false,"type":"enum","values":["zero_shot","cross_lingual","instruct"]},{"name":"reference_text","required":false,"type":"string"},{"name":"instruction","required":false,"type":"string"},{"default":600,"min":1,"name":"text_chunk_size","required":false,"type":"int"},{"default":"default","name":"text_chunk_mode","required":false,"type":"enum","values":["default","tag_aware","japanese","endline"]},{"default":1600,"min":1,"name":"max_tokens","required":false,"type":"int"},{"default":0,"min":0,"name":"min_tokens","required":false,"type":"int"},{"default":25,"min":1,"name":"top_k","required":false,"type":"int"},{"default":10,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"default":1986,"min":0,"name":"seed","required":false,"type":"int"}]},"schema_version":1,"tasks":["tts","clone"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/DotTTS-Edit-GGUF": {"family":"dots_tts","files":["dots-tts-edit-bf16.gguf","dots-tts-edit-q8_0.gguf"],"spec":{"options":{"request":[{"name":"reference_text","required":false,"type":"string"},{"min":0,"name":"reference_duration_sec","required":false,"type":"float"},{"default":"tts","name":"template_name","required":false,"type":"enum","values":["tts","instruction_tts","text_to_audio","tts_interleave","edit"]},{"name":"source_audio","required":false,"type":"audio_path"},{"name":"instruction","required":false,"type":"string"},{"name":"instruct","required":false,"type":"string"},{"name":"source_text","required":false,"type":"string"},{"name":"target_text","required":false,"type":"string"},{"default":"auto","name":"use_xvector","required":false,"type":"enum","values":["auto","on","off"]},{"default":"none","name":"language","required":false,"type":"string"},{"default":10,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"default":1.2,"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":1.5,"min":0,"name":"speaker_scale","required":false,"type":"float"},{"default":"euler","name":"sampler_mode","required":false,"type":"enum","values":["euler","midpoint","rk4"]},{"default":500,"min":1,"name":"max_tokens","required":false,"type":"int"},{"default":320,"min":1,"name":"text_chunk_size","required":false,"type":"int"},{"default":"tag_aware","name":"text_chunk_mode","required":false,"type":"enum","values":["default","tag_aware","japanese","endline"]},{"default":4,"min":1,"name":"vocoder_merge_steps","required":false,"type":"int"},{"default":42,"min":0,"name":"seed","required":false,"type":"int"}]},"schema_version":1,"tasks":["tts","clone"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/DotTTS-MF-GGUF": {"family":"dots_tts","files":["dots-tts-mf-bf16.gguf","dots-tts-mf-q8_0.gguf"],"spec":{"options":{"request":[{"name":"reference_text","required":false,"type":"string"},{"min":0,"name":"reference_duration_sec","required":false,"type":"float"},{"default":"tts","name":"template_name","required":false,"type":"enum","values":["tts","instruction_tts","text_to_audio","tts_interleave"]},{"default":"none","name":"language","required":false,"type":"string"},{"default":10,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"default":1.2,"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":1.5,"min":0,"name":"speaker_scale","required":false,"type":"float"},{"default":"euler","name":"sampler_mode","required":false,"type":"enum","values":["euler","midpoint","rk4"]},{"default":500,"min":1,"name":"max_tokens","required":false,"type":"int"},{"default":320,"min":1,"name":"text_chunk_size","required":false,"type":"int"},{"default":"tag_aware","name":"text_chunk_mode","required":false,"type":"enum","values":["default","tag_aware","japanese","endline"]},{"default":4,"min":1,"name":"vocoder_merge_steps","required":false,"type":"int"},{"default":42,"min":0,"name":"seed","required":false,"type":"int"}]},"schema_version":1,"tasks":["tts","clone"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/DotTTS-SOAR-GGUF": {"family":"dots_tts","files":["dots-tts-soar-bf16.gguf","dots-tts-soar-q8_0.gguf"],"spec":{"options":{"request":[{"name":"reference_text","required":false,"type":"string"},{"min":0,"name":"reference_duration_sec","required":false,"type":"float"},{"default":"tts","name":"template_name","required":false,"type":"enum","values":["tts","instruction_tts","text_to_audio","tts_interleave"]},{"default":"none","name":"language","required":false,"type":"string"},{"default":10,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"default":1.2,"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":1.5,"min":0,"name":"speaker_scale","required":false,"type":"float"},{"default":"euler","name":"sampler_mode","required":false,"type":"enum","values":["euler","midpoint","rk4"]},{"default":500,"min":1,"name":"max_tokens","required":false,"type":"int"},{"default":320,"min":1,"name":"text_chunk_size","required":false,"type":"int"},{"default":"tag_aware","name":"text_chunk_mode","required":false,"type":"enum","values":["default","tag_aware","japanese","endline"]},{"default":4,"min":1,"name":"vocoder_merge_steps","required":false,"type":"int"},{"default":42,"min":0,"name":"seed","required":false,"type":"int"}]},"schema_version":1,"tasks":["tts","clone"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/DramaBox-GGUF": {"family":"dramabox","files":["dramabox-q8_0.gguf"],"spec":{"options":{"request":[{"name":"target_voice","required":false,"type":"audio_path"},{"name":"negative_prompt","required":false,"type":"string"},{"default":0,"min":0,"name":"duration_sec","required":false,"type":"float"},{"default":30,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"default":2.5,"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":1.5,"min":0,"name":"spatio_temporal_guidance_scale","required":false,"type":"float"},{"default":1.1,"min":0,"name":"duration_scale","required":false,"type":"float"},{"default":10,"min":0,"name":"reference_duration_sec","required":false,"type":"float"},{"default":"auto","name":"guidance_rescale","required":false,"type":"string"},{"default":45,"min":0,"name":"audio_chunk_threshold_sec","required":false,"type":"float"},{"default":37,"min":0,"name":"audio_chunk_duration_sec","required":false,"type":"float"},{"default":0.05,"min":0,"name":"cross_fade_duration_sec","required":false,"type":"float"},{"default":42,"min":0,"name":"seed","required":false,"type":"int"}]},"schema_version":1,"tasks":["tts","clone"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/FireRedAudio-GGUF": {"family":"firered_audio","files":["firered-audio-orig.gguf","firered-audio-q8_0.gguf"],"spec":{"options":{"request":[{"default":"tts_clone","name":"template_name","required":false,"type":"enum","values":["tts_clone","voice_design","semantic_edit","acoustic_edit","asr","understand"]},{"name":"reference_text","required":false,"type":"string"},{"name":"instruction","required":false,"type":"string"},{"default":"zh","name":"language","required":false,"type":"string"},{"default":10,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"default":2,"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":750,"min":1,"name":"max_new_audio_steps","required":false,"type":"int"},{"default":6,"min":0,"name":"min_new_audio_steps","required":false,"type":"int"},{"default":512,"min":1,"name":"max_new_text_tokens","required":false,"type":"int"},{"default":300,"min":1,"name":"max_new_tokens","required":false,"type":"int"},{"default":false,"name":"enable_thinking","required":false,"type":"bool"},{"default":20,"min":0,"name":"top_k","required":false,"type":"int"},{"default":0.8,"max":1,"min":0,"name":"top_p","required":false,"type":"float"},{"default":0.7,"min":0,"name":"temperature","required":false,"type":"float"},{"default":1234,"min":0,"name":"seed","required":false,"type":"int"}]},"schema_version":1,"tasks":["asr","tts","clone","design"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/FireRedTTS3-Base-GGUF": {"family":"fireredtts3","files":["fireredtts3-base-orig.gguf","fireredtts3-base-q8_0.gguf"],"spec":{"options":{"request":[{"default":"instruct_tts","name":"template_name","required":false,"type":"enum","values":["instruct_tts","voice_design","semantic_edit","acoustic_edit"]},{"name":"reference_text","required":false,"type":"string"},{"name":"instruction","required":false,"type":"string"},{"default":"Chinese","name":"language","required":false,"type":"string"},{"default":600,"min":1,"name":"text_chunk_size","required":false,"type":"int"},{"default":"default","name":"text_chunk_mode","required":false,"type":"enum","values":["default","tag_aware","japanese","endline"]},{"default":10,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"default":2,"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":0.5,"max":1,"min":0,"name":"stop_threshold","required":false,"type":"float"},{"default":1234,"min":0,"name":"seed","required":false,"type":"int"}]},"schema_version":1,"tasks":["tts","clone","design"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/FireRedTTS3-Instruct-GGUF": {"family":"fireredtts3","files":["fireredtts3-instruct-orig.gguf","fireredtts3-instruct-q8_0.gguf"],"spec":{"options":{"request":[{"default":"instruct_tts","name":"template_name","required":false,"type":"enum","values":["instruct_tts","voice_design","semantic_edit","acoustic_edit"]},{"name":"reference_text","required":false,"type":"string"},{"name":"instruction","required":false,"type":"string"},{"default":"Chinese","name":"language","required":false,"type":"string"},{"default":600,"min":1,"name":"text_chunk_size","required":false,"type":"int"},{"default":"default","name":"text_chunk_mode","required":false,"type":"enum","values":["default","tag_aware","japanese","endline"]},{"default":10,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"default":2,"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":0.5,"max":1,"min":0,"name":"stop_threshold","required":false,"type":"float"},{"default":1234,"min":0,"name":"seed","required":false,"type":"int"}]},"schema_version":1,"tasks":["tts","clone","design"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Fish-Audio-S2-Pro-GGUF": {"family":"fish_audio","files":["fish-audio-s2-pro-bf16.gguf","fish-audio-s2-pro-q8_0.gguf"],"spec":{},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/HTDemucs-6stems-GGUF": {"family":"htdemucs_6stems","files":["htdemucs-6stems-f16.gguf","htdemucs-6stems-q8_0.gguf"],"spec":{"schema_version":1,"tasks":["sep"]},"task":"sep"},
+ "audio-cpp/audio.cpp-gguf/HTDemucs-GGUF": {"family":"htdemucs","files":["htdemucs-f16.gguf","htdemucs-q8_0.gguf"],"spec":{"schema_version":1,"tasks":["sep"]},"task":"sep"},
+ "audio-cpp/audio.cpp-gguf/HeartMuLa-GGUF": {"family":"heartmula","files":["heartmula-f16.gguf","heartmula-q8_0.gguf"],"spec":{"options":{"request":[{"name":"lyrics","required":false,"type":"string"},{"name":"tags","required":false,"type":"string"},{"default":120,"min":0,"name":"duration_sec","required":false,"type":"float"},{"default":1,"min":0,"name":"temperature","required":false,"type":"float"},{"default":50,"min":1,"name":"top_k","required":false,"type":"int"},{"default":1.5,"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":29.76,"min":0,"name":"codec_duration_sec","required":false,"type":"float"},{"default":10,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"default":1.25,"min":0,"name":"codec_guidance_scale","required":false,"type":"float"},{"default":false,"name":"infinite_mode","required":false,"type":"bool"},{"default":4096,"min":1,"name":"text_chunk_size","required":false,"type":"int"},{"default":240000,"min":1,"name":"infinite_chunk_audio_duration_ms","required":false,"type":"int"},{"default":1234,"min":0,"name":"seed","required":false,"type":"int"}]},"schema_version":1,"tasks":["music"]},"task":"music"},
+ "audio-cpp/audio.cpp-gguf/Higgs-Audio-v3-TTS-4B-GGUF": {"family":"higgs_audio_tts","files":["higgs-audio-v3-tts-4b-bf16.gguf","higgs-audio-v3-tts-4b-q8_0.gguf"],"spec":{},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/IndexTTS2-GGUF": {"family":"index_tts2","files":["index-tts2-f16.gguf","index-tts2-orig.gguf","index-tts2-q8_0.gguf"],"spec":{},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/IndexTTS2.5-GGUF": {"family":"index_tts2","files":["index-tts2_5-f16.gguf","index-tts2_5-orig.gguf","index-tts2_5-q8_0.gguf"],"spec":{"tasks":["tts","clone"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Inflect-Micro-v2-GGUF": {"family":"inflect_v2","files":["inflect-micro-v2-orig.gguf"],"spec":{"options":{"request":[{"default":1,"max":2,"min":0.5,"name":"speaking_rate","required":false,"type":"float"},{"default":0.667,"max":1,"min":0,"name":"variation","required":false,"type":"float"},{"default":0,"min":0,"name":"seed","required":false,"type":"int"},{"default":"word_budget","name":"text_chunk_mode","required":false,"type":"enum","values":["word_budget"]},{"default":280,"min":1,"name":"text_chunk_size","required":false,"type":"int"}]},"schema_version":1,"tasks":["tts"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Irodori-TTS-500M-v3-GGUF": {"family":"irodori_tts","files":["irodori-tts-500m-v3-f16.gguf","irodori-tts-500m-v3-q8_0.gguf"],"spec":{"options":{"request":[{"default":"ja","name":"language","required":false,"type":"string"},{"name":"caption","required":false,"type":"string"},{"default":true,"name":"no_ref","required":false,"type":"bool"},{"default":40,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"min":0,"name":"duration_sec","required":false,"type":"float"},{"default":1,"min":0,"name":"duration_scale","required":false,"type":"float"},{"default":"endline","name":"text_chunk_mode","required":false,"type":"enum","values":["japanese","endline"]},{"min":1,"name":"text_chunk_size","required":false,"type":"int"},{"default":0.5,"min":0,"name":"min_duration_sec","required":false,"type":"float"},{"default":30,"min":0,"name":"max_duration_sec","required":false,"type":"float"},{"default":3,"min":0,"name":"text_guidance_scale","required":false,"type":"float"},{"default":5,"min":0,"name":"speaker_guidance_scale","required":false,"type":"float"},{"default":3,"min":0,"name":"caption_guidance_scale","required":false,"type":"float"},{"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":"independent","name":"guidance_mode","required":false,"type":"enum","values":["independent","joint","alternating"]},{"default":0.5,"min":0,"name":"guidance_min_t","required":false,"type":"float"},{"default":1,"min":0,"name":"guidance_max_t","required":false,"type":"float"},{"min":0,"name":"seed","required":false,"type":"int"},{"default":true,"name":"trim_tail","required":false,"type":"bool"}]},"schema_version":1,"tasks":["tts","clone","design"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Irodori-TTS-600M-v3-VoiceDesign-GGUF": {"family":"irodori_tts","files":["irodori-tts-600m-v3-voicedesign-f16.gguf","irodori-tts-600m-v3-voicedesign-q8_0.gguf"],"spec":{"options":{"request":[{"default":"ja","name":"language","required":false,"type":"string"},{"name":"caption","required":false,"type":"string"},{"default":true,"name":"no_ref","required":false,"type":"bool"},{"default":40,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"min":0,"name":"duration_sec","required":false,"type":"float"},{"default":1,"min":0,"name":"duration_scale","required":false,"type":"float"},{"default":"endline","name":"text_chunk_mode","required":false,"type":"enum","values":["japanese","endline"]},{"min":1,"name":"text_chunk_size","required":false,"type":"int"},{"default":0.5,"min":0,"name":"min_duration_sec","required":false,"type":"float"},{"default":30,"min":0,"name":"max_duration_sec","required":false,"type":"float"},{"default":3,"min":0,"name":"text_guidance_scale","required":false,"type":"float"},{"default":5,"min":0,"name":"speaker_guidance_scale","required":false,"type":"float"},{"default":3,"min":0,"name":"caption_guidance_scale","required":false,"type":"float"},{"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":"independent","name":"guidance_mode","required":false,"type":"enum","values":["independent","joint","alternating"]},{"default":0.5,"min":0,"name":"guidance_min_t","required":false,"type":"float"},{"default":1,"min":0,"name":"guidance_max_t","required":false,"type":"float"},{"min":0,"name":"seed","required":false,"type":"int"},{"default":true,"name":"trim_tail","required":false,"type":"bool"}]},"schema_version":1,"tasks":["tts","clone","design"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Irodori-TTS-v4-Small-GGUF": {"family":"irodori_tts","files":["irodori-tts-v4-small-f16.gguf","irodori-tts-v4-small-q8_0.gguf","irodori-tts-v4.1-anime-q8_0.gguf"],"spec":{"options":{"request":[{"default":"ja","name":"language","required":false,"type":"string"},{"name":"instruction","required":false,"type":"string"},{"default":true,"name":"no_ref","required":false,"type":"bool"},{"default":40,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"min":0,"name":"duration_sec","required":false,"type":"float"},{"default":1,"min":0,"name":"duration_scale","required":false,"type":"float"},{"default":"endline","name":"text_chunk_mode","required":false,"type":"enum","values":["japanese","endline"]},{"min":1,"name":"text_chunk_size","required":false,"type":"int"},{"default":0.5,"min":0,"name":"min_duration_sec","required":false,"type":"float"},{"default":30,"min":0,"name":"max_duration_sec","required":false,"type":"float"},{"default":3,"min":0,"name":"text_guidance_scale","required":false,"type":"float"},{"default":5,"min":0,"name":"speaker_guidance_scale","required":false,"type":"float"},{"default":3,"min":0,"name":"caption_guidance_scale","required":false,"type":"float"},{"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":"independent","name":"guidance_mode","required":false,"type":"enum","values":["independent","joint","alternating"]},{"default":0.5,"min":0,"name":"guidance_min_t","required":false,"type":"float"},{"default":1,"min":0,"name":"guidance_max_t","required":false,"type":"float"},{"min":0,"name":"seed","required":false,"type":"int"},{"default":true,"name":"trim_tail","required":false,"type":"bool"}]},"schema_version":1,"tasks":["tts","clone","design"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/KittenTTS-GGUF": {"family":"kitten_tts","files":["LICENSE","README.md","kitten-tts-mini-0.8-orig.gguf"],"spec":{"options":{"request":[{"default":1,"min":0.01,"name":"speed","required":false,"type":"float"},{"min":0,"name":"seed","required":false,"type":"int"},{"default":"default","name":"text_chunk_mode","required":false,"type":"enum","values":["default","tag_aware","japanese","endline"]},{"default":400,"min":32,"name":"text_chunk_size","required":false,"type":"int"}]},"schema_version":1,"tasks":["tts"],"ui":{"builtin_voices":["Bella","Jasper","Luna","Bruno","Rosie","Hugo","Kiki","Leo"],"default_voice":"Leo"}},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Kokoro-82M-GGUF": {"family":"kokoro_tts","files":["kokoro-82m-bf16.gguf","kokoro-82m-q8_0.gguf"],"spec":{"options":{"request":[{"name":"language","required":false,"type":"string"},{"min":0,"name":"seed","required":false,"type":"int"},{"default":1,"min":0.01,"name":"speed","required":false,"type":"float"},{"name":"phonemes","required":false,"type":"string_list"},{"default":240,"min":32,"name":"text_chunk_size","required":false,"type":"int"},{"default":false,"name":"return_timestamps","required":false,"type":"bool"}]},"schema_version":1,"tasks":["tts"],"ui":{"builtin_voices":["af_alloy","af_aoede","af_bella","af_heart","af_jessica","af_kore","af_nicole","af_nova","af_river","af_sarah","af_sky","am_adam","am_echo","am_eric","am_fenrir","am_liam","am_michael","am_onyx","am_puck","am_santa","bf_alice","bf_emma","bf_isabella","bf_lily","bm_daniel","bm_fable","bm_george","bm_lewis","ef_dora","em_alex","em_santa","ff_siwis","hf_alpha","hf_beta","hm_omega","hm_psi","if_sara","im_nicola","jf_alpha","jf_gongitsune","jf_nezumi","jf_tebukuro","jm_kumo","pf_dora","pm_alex","pm_santa","zf_xiaobei","zf_xiaoni","zf_xiaoxiao","zf_xiaoyi","zm_yunjian","zm_yunxi","zm_yunxia","zm_yunyang"],"default_voice":"af_heart"}},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/MOSS-TTS-Local-v1.5-GGUF": {"family":"moss_tts_local","files":["moss-tts-local-v1.5-bf16.gguf","moss-tts-local-v1.5-q8_0.gguf"],"spec":{},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/MOSS-TTS-Nano-100M-GGUF": {"family":"moss_tts_nano","files":["moss-tts-nano-100m-bf16.gguf","moss-tts-nano-100m-q8_0.gguf"],"spec":{},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/MOSS-VoiceGenerator-GGUF": {"family":"moss_voicegen","files":["moss_voicegen_bf16_codec_f16_decode.gguf"],"spec":{"tasks":["vdes"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/MagpieTTS-Multilingual-357M-GGUF": {"family":"magpie_tts","files":["magpie-tts-multilingual-357m-orig.gguf","magpie-tts-multilingual-357m-q8_0.gguf"],"spec":{"options":{"request":[{"default":"en","name":"language","required":false,"type":"string"},{"default":"0","name":"voice_id","required":false,"type":"enum","values":["0","Sofia","Jason"]},{"default":0.6,"min":0,"name":"temperature","required":false,"type":"float"},{"default":80,"min":1,"name":"top_k","required":false,"type":"int"},{"default":2.5,"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":500,"min":1,"name":"max_tokens","required":false,"type":"int"},{"default":300,"min":1,"name":"text_chunk_size","required":false,"type":"int"},{"default":"default","name":"text_chunk_mode","required":false,"type":"enum","values":["default","tag_aware","japanese","endline"]},{"default":0,"min":0,"name":"seed","required":false,"type":"int"}]},"schema_version":1,"tasks":["tts"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Maya1-GGUF": {"family":"maya1","files":["LICENSE-MAYA1","LICENSE-SNAC","README.md","maya1-orig.gguf","maya1-q8_0.gguf"],"spec":{"options":{"request":[{"name":"instruct","required":true,"type":"string"},{"default":2048,"min":1,"name":"max_tokens","required":false,"type":"int"},{"default":28,"min":0,"name":"min_tokens","required":false,"type":"int"},{"default":0.4,"min":0,"name":"temperature","required":false,"type":"float"},{"default":0.9,"max":1,"min":0,"name":"top_p","required":false,"type":"float"},{"default":1.1,"min":0,"name":"repetition_penalty","required":false,"type":"float"},{"min":0,"name":"seed","required":false,"type":"int"},{"default":300,"min":1,"name":"text_chunk_size","required":false,"type":"int"},{"default":"tag_aware","name":"text_chunk_mode","required":false,"type":"enum","values":["default","tag_aware","japanese","endline"]}]},"schema_version":1,"tasks":["tts"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/MeanVC2-GGUF": {"family":"meanvc2","files":["meanvc2-120ms-40ms-fp32.gguf","meanvc2-120ms-40ms-q4_k.gguf"],"spec":{"options":{"request":[{"default":42,"min":0,"name":"seed","required":false,"type":"int"}]},"schema_version":1,"tasks":["vc"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Mel-Band-RoFormer-GGUF": {"family":"mel_band_roformer","files":["mel-band-roformer-f16.gguf","mel-band-roformer-q8_0.gguf"],"spec":{"schema_version":1,"tasks":["sep"]},"task":"sep"},
+ "audio-cpp/audio.cpp-gguf/MiDashengLM-Gen-GGUF": {"family":"midashenglm_gen","files":["midashenglm-gen-f32.gguf","midashenglm-gen-q8_0.gguf"],"spec":{"options":{"request":[{"default":20,"max":163.84,"min":0.04,"name":"duration_sec","required":false,"type":"float"},{"default":2,"max":20,"min":0,"name":"guidance_scale","required":false,"type":"float"},{"default":0.5,"max":1,"min":0,"name":"stop_threshold","required":false,"type":"float"},{"default":5,"max":4096,"min":0,"name":"min_stop_step","required":false,"type":"int"},{"default":0,"max":2147483647,"min":-1,"name":"seed","required":false,"type":"int"}]},"schema_version":1,"tasks":["music"]},"task":"music"},
+ "audio-cpp/audio.cpp-gguf/MioTTS-1.7B-GGUF": {"family":"miotts","files":["miotts-1.7b-bf16.gguf","miotts-1.7b-orig.gguf","miotts-1.7b-q8_0.gguf"],"spec":{},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/NeuTTS-2E-GGUF": {"family":"neutts","files":["neutts-2e-orig.gguf"],"spec":{"options":{"request":[{"default":"emily","name":"voice_id","required":false,"type":"enum","values":["dave","emily","greta","jo","juliette","mateo","paul","sophie","steven"]},{"default":"neutral","name":"emotion","required":false,"type":"enum","values":["angry","disgusted","sad","happy","fearful","neutral","surprised"]},{"default":0,"min":0,"name":"max_tokens","required":false,"type":"int"},{"default":50,"min":0,"name":"min_tokens","required":false,"type":"int"},{"default":1,"min":0,"name":"temperature","required":false,"type":"float"},{"default":50,"min":1,"name":"top_k","required":false,"type":"int"},{"min":0,"name":"seed","required":false,"type":"int"},{"default":"default","name":"text_chunk_mode","required":false,"type":"enum","values":["default","tag_aware","japanese","endline"]},{"default":600,"min":1,"name":"text_chunk_size","required":false,"type":"int"}]},"schema_version":1,"tasks":["tts"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/OmniVoice-GGUF": {"family":"omnivoice","files":["omnivoice-bf16.gguf","omnivoice-f16.gguf","omnivoice-q8_0.gguf"],"spec":{},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Piper-TTS-GGUF": {"family":"piper_tts","files":["LICENSE","README.md","piper-en-us-lessac-medium-orig.gguf"],"spec":{"options":{"request":[{"default":1,"max":2,"min":0.5,"name":"speed","required":false,"type":"float"},{"default":0.667,"max":1,"min":0,"name":"variation","required":false,"type":"float"},{"default":0.8,"max":2,"min":0,"name":"duration_variation","required":false,"type":"float"},{"default":1234,"min":0,"name":"seed","required":false,"type":"int"},{"default":"default","name":"text_chunk_mode","required":false,"type":"enum","values":["default","tag_aware","japanese","endline"]},{"default":280,"min":1,"name":"text_chunk_size","required":false,"type":"int"}]},"schema_version":1,"tasks":["tts"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/PocketTTS-GGUF": {"family":"pocket_tts","files":["english/embeddings/alba.safetensors","english/embeddings/anna.safetensors","english/embeddings/azelma.safetensors","english/embeddings/bill_boerst.safetensors","english/embeddings/caro_davy.safetensors","english/embeddings/charles.safetensors","english/embeddings/cosette.safetensors","english/embeddings/eponine.safetensors","english/embeddings/estelle.safetensors","english/embeddings/eve.safetensors","english/embeddings/fantine.safetensors","english/embeddings/george.safetensors","english/embeddings/giovanni.safetensors","english/embeddings/jane.safetensors","english/embeddings/javert.safetensors","english/embeddings/jean.safetensors","english/embeddings/juergen.safetensors","english/embeddings/lola.safetensors","english/embeddings/marius.safetensors","english/embeddings/mary.safetensors","english/embeddings/michael.safetensors","english/embeddings/paul.safetensors","english/embeddings/peter_yearsley.safetensors","english/embeddings/rafael.safetensors","english/embeddings/stuart_bell.safetensors","english/embeddings/vera.safetensors","english/pocket-tts-english-bf16.gguf","english/pocket-tts-english-q8_0.gguf","french/embeddings/alba.safetensors","french/embeddings/anna.safetensors","french/embeddings/azelma.safetensors","french/embeddings/bill_boerst.safetensors","french/embeddings/caro_davy.safetensors","french/embeddings/charles.safetensors","french/embeddings/cosette.safetensors","french/embeddings/daan.safetensors","french/embeddings/eponine.safetensors","french/embeddings/estelle.safetensors","french/embeddings/eve.safetensors","french/embeddings/fantine.safetensors","french/embeddings/george.safetensors","french/embeddings/giovanni.safetensors","french/embeddings/jane.safetensors","french/embeddings/javert.safetensors","french/embeddings/jean.safetensors","french/embeddings/juergen.safetensors","french/embeddings/lola.safetensors","french/embeddings/marius.safetensors","french/embeddings/mary.safetensors","french/embeddings/michael.safetensors","french/embeddings/paul.safetensors","french/embeddings/peter_yearsley.safetensors","french/embeddings/rafael.safetensors","french/embeddings/stuart_bell.safetensors","french/embeddings/vera.safetensors","french/pocket-tts-french-bf16.gguf","french/pocket-tts-french-q8_0.gguf","german/embeddings/alba.safetensors","german/embeddings/anna.safetensors","german/embeddings/azelma.safetensors","german/embeddings/bill_boerst.safetensors","german/embeddings/caro_davy.safetensors","german/embeddings/charles.safetensors","german/embeddings/cosette.safetensors","german/embeddings/eponine.safetensors","german/embeddings/estelle.safetensors","german/embeddings/eve.safetensors","german/embeddings/fantine.safetensors","german/embeddings/george.safetensors","german/embeddings/giovanni.safetensors","german/embeddings/jane.safetensors","german/embeddings/javert.safetensors","german/embeddings/jean.safetensors","german/embeddings/juergen.safetensors","german/embeddings/lola.safetensors","german/embeddings/marius.safetensors","german/embeddings/mary.safetensors","german/embeddings/michael.safetensors","german/embeddings/paul.safetensors","german/embeddings/peter_yearsley.safetensors","german/embeddings/rafael.safetensors","german/embeddings/stuart_bell.safetensors","german/embeddings/vera.safetensors","german/pocket-tts-german-bf16.gguf","german/pocket-tts-german-q8_0.gguf","italian/embeddings/alba.safetensors","italian/embeddings/anna.safetensors","italian/embeddings/azelma.safetensors","italian/embeddings/bill_boerst.safetensors","italian/embeddings/caro_davy.safetensors","italian/embeddings/charles.safetensors","italian/embeddings/cosette.safetensors","italian/embeddings/eponine.safetensors","italian/embeddings/estelle.safetensors","italian/embeddings/eve.safetensors","italian/embeddings/fantine.safetensors","italian/embeddings/george.safetensors","italian/embeddings/giovanni.safetensors","italian/embeddings/jane.safetensors","italian/embeddings/javert.safetensors","italian/embeddings/jean.safetensors","italian/embeddings/juergen.safetensors","italian/embeddings/lola.safetensors","italian/embeddings/marius.safetensors","italian/embeddings/mary.safetensors","italian/embeddings/michael.safetensors","italian/embeddings/paul.safetensors","italian/embeddings/peter_yearsley.safetensors","italian/embeddings/rafael.safetensors","italian/embeddings/stuart_bell.safetensors","italian/embeddings/vera.safetensors","italian/pocket-tts-italian-bf16.gguf","italian/pocket-tts-italian-q8_0.gguf","portuguese/embeddings/alba.safetensors","portuguese/embeddings/anna.safetensors","portuguese/embeddings/azelma.safetensors","portuguese/embeddings/bill_boerst.safetensors","portuguese/embeddings/caro_davy.safetensors","portuguese/embeddings/charles.safetensors","portuguese/embeddings/cosette.safetensors","portuguese/embeddings/eponine.safetensors","portuguese/embeddings/estelle.safetensors","portuguese/embeddings/eve.safetensors","portuguese/embeddings/fantine.safetensors","portuguese/embeddings/george.safetensors","portuguese/embeddings/giovanni.safetensors","portuguese/embeddings/jane.safetensors","portuguese/embeddings/javert.safetensors","portuguese/embeddings/jean.safetensors","portuguese/embeddings/juergen.safetensors","portuguese/embeddings/lola.safetensors","portuguese/embeddings/marius.safetensors","portuguese/embeddings/mary.safetensors","portuguese/embeddings/michael.safetensors","portuguese/embeddings/paul.safetensors","portuguese/embeddings/peter_yearsley.safetensors","portuguese/embeddings/rafael.safetensors","portuguese/embeddings/stuart_bell.safetensors","portuguese/embeddings/vera.safetensors","portuguese/pocket-tts-portuguese-bf16.gguf","portuguese/pocket-tts-portuguese-q8_0.gguf","spanish/embeddings/alba.safetensors","spanish/embeddings/anna.safetensors","spanish/embeddings/azelma.safetensors","spanish/embeddings/bill_boerst.safetensors","spanish/embeddings/caro_davy.safetensors","spanish/embeddings/charles.safetensors","spanish/embeddings/cosette.safetensors","spanish/embeddings/eponine.safetensors","spanish/embeddings/estelle.safetensors","spanish/embeddings/eve.safetensors","spanish/embeddings/fantine.safetensors","spanish/embeddings/george.safetensors","spanish/embeddings/giovanni.safetensors","spanish/embeddings/jane.safetensors","spanish/embeddings/javert.safetensors","spanish/embeddings/jean.safetensors","spanish/embeddings/juergen.safetensors","spanish/embeddings/lola.safetensors","spanish/embeddings/marius.safetensors","spanish/embeddings/mary.safetensors","spanish/embeddings/michael.safetensors","spanish/embeddings/paul.safetensors","spanish/embeddings/peter_yearsley.safetensors","spanish/embeddings/rafael.safetensors","spanish/embeddings/stuart_bell.safetensors","spanish/embeddings/vera.safetensors","spanish/pocket-tts-spanish-bf16.gguf","spanish/pocket-tts-spanish-q8_0.gguf"],"spec":{},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Qwen3-TTS-12Hz-0.6B-Base-GGUF": {"family":"qwen3_tts","files":["qwen3-tts-12hz-0.6b-base-bf16.gguf","qwen3-tts-12hz-0.6b-base-q8_0.gguf"],"spec":{"tasks":["tts","clone","design"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Qwen3-TTS-12Hz-1.7B-Base-GGUF": {"family":"qwen3_tts","files":["qwen3-tts-12hz-1.7b-base-bf16.gguf","qwen3-tts-12hz-1.7b-base-orig.gguf","qwen3-tts-12hz-1.7b-base-q8_0_v2.gguf"],"spec":{},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Qwen3-TTS-12Hz-1.7B-CustomVoice-GGUF": {"family":"qwen3_tts","files":["qwen3-tts-12hz-1.7b-customvoice-bf16.gguf","qwen3-tts-12hz-1.7b-customvoice-q8_0.gguf"],"spec":{},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Qwen3-TTS-12Hz-1.7B-VoiceDesign-GGUF": {"family":"qwen3_tts","files":["qwen3-tts-12hz-1.7b-voicedesign-bf16.gguf","qwen3-tts-12hz-1.7b-voicedesign-q8_0.gguf"],"spec":{},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/RVC-GGUF": {"family":"rvc","files":["rvc-f16.gguf"],"spec":{"options":{"request":[{"default":"default","name":"voice_id","required":false,"type":"enum","values":["default","manthos","chocola","fraise"]},{"name":"voice_model_path","required":false,"type":"path"},{"default":"rmvpe","name":"pitch_extractor","required":false,"type":"enum","values":["rmvpe"]},{"name":"pitch_path","required":false,"type":"path"},{"name":"retrieval_index_path","required":false,"type":"path"},{"default":0,"max":1,"min":0,"name":"retrieval_blend","required":false,"type":"float"},{"default":0,"name":"semitone_shift","required":false,"type":"int"},{"default":3,"min":0,"name":"pitch_filter_radius","required":false,"type":"int"},{"default":0,"min":0,"name":"output_sample_rate","required":false,"type":"int"},{"default":0.25,"name":"rms_mix_rate","required":false,"type":"float"},{"default":0.33,"max":1,"min":0,"name":"unvoiced_protection","required":false,"type":"float"},{"default":0,"min":0,"name":"speaker_id","required":false,"type":"int"},{"default":1,"min":1,"name":"audio_pad_duration_sec","required":false,"type":"int"},{"default":5,"min":1,"name":"split_query_sec","required":false,"type":"int"},{"default":30,"min":1,"name":"split_center_sec","required":false,"type":"int"},{"default":32,"min":1,"name":"split_threshold_sec","required":false,"type":"int"}]},"schema_version":1,"tasks":["vc"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/SeedVC-MLX-GGUF": {"family":"seed_vc","files":["seed-vc-mlx-f16.gguf","seed-vc-mlx-orig.gguf","seed-vc-mlx-q8_0.gguf"],"spec":{"options":{"request":[{"name":"route","required":false,"type":"enum","values":["v2_vc","v1_svc","v1_whisper_bigvgan_vc","v1_xlsr_hift_vc"]},{"default":1,"min":0,"name":"length_adjust","required":false,"type":"float"},{"default":30,"min":1,"name":"num_inference_steps","required":false,"type":"int"},{"default":0.7,"min":0,"name":"inference_guidance_scale","required":false,"type":"float"},{"default":0.7,"min":0,"name":"intelligibility_guidance_scale","required":false,"type":"float"},{"default":0.7,"min":0,"name":"similarity_guidance_scale","required":false,"type":"float"},{"default":false,"name":"voice_anonymization","required":false,"type":"bool"},{"min":0,"name":"seed","required":false,"type":"int"},{"name":"noise_path","required":false,"type":"path"},{"default":false,"name":"f0_condition","required":false,"type":"bool"},{"default":false,"name":"auto_f0_adjust","required":false,"type":"bool"},{"default":0,"name":"semitone_shift","required":false,"type":"int"}]},"schema_version":1,"tasks":["vc","svc"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Stable-Audio-3-Medium-GGUF": {"family":"stable_audio","files":["stable-audio-3-medium-f16.gguf","stable-audio-3-medium-q8_0.gguf"],"spec":{},"task":"music"},
+ "audio-cpp/audio.cpp-gguf/Stable-Audio-3-Small-Music-GGUF": {"family":"stable_audio","files":["stable-audio-3-small-music-f16.gguf","stable-audio-3-small-music-q8_0.gguf"],"spec":{},"task":"music"},
+ "audio-cpp/audio.cpp-gguf/Stable-Audio-3-Small-SFX-GGUF": {"family":"stable_audio","files":["stable-audio-3-small-sfx-f16.gguf","stable-audio-3-small-sfx-q8_0.gguf"],"spec":{},"task":"music"},
+ "audio-cpp/audio.cpp-gguf/Supertonic-3-GGUF": {"family":"supertonic","files":["supertonic-3-f16.gguf","supertonic-3-orig.gguf","supertonic-3-q8_0.gguf"],"spec":{},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Tone-Color-VC-GGUF": {"family":"tone_color_vc","files":["LICENSE","tone-color-vc-f16.gguf","tone-color-vc-f32.gguf"],"spec":{"options":{"request":[{"default":1234,"min":-1,"name":"seed","required":false,"type":"int"},{"default":0.3,"min":0,"name":"temperature","required":false,"type":"float"}]},"schema_version":1,"tasks":["vc"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/Vevo2-GGUF": {"family":"vevo2","files":["vevo2-f16.gguf","vevo2-orig.gguf","vevo2-q8_0.gguf"],"spec":{},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/VibeVoice-1.5B-GGUF": {"family":"vibevoice","files":["vibevoice-1.5b-bf16.gguf","vibevoice-1.5b-q4-ios.gguf","vibevoice-1.5b-q8_0.gguf"],"spec":{},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/VoxCPM1-GGUF": {"family":"voxcpm1","files":["voxcpm-0.5b-q8_0-audiovae-f16.gguf"],"spec":{"options":{"request":[{"default":"tag_aware","name":"text_chunk_mode","preset":"text_chunk_mode_full","required":false,"type":"enum"},{"name":"seed","required":false,"type":"int"},{"default":1024,"name":"max_tokens","required":false,"type":"int"},{"default":0,"name":"min_tokens","required":false,"type":"int"},{"default":50,"name":"num_inference_steps","required":false,"type":"int"},{"default":2,"name":"guidance_scale","required":false,"type":"float"},{"default":true,"name":"retry_badcase","required":false,"type":"bool"},{"default":2,"name":"retry_badcase_max_times","required":false,"type":"int"},{"name":"retry_badcase_ratio_threshold","required":false,"type":"float"},{"name":"prompt_text","required":false,"type":"string"},{"name":"reference_text","required":false,"type":"string"}]},"schema_version":1,"tasks":["tts","clone"]},"task":"tts"},
+ "audio-cpp/audio.cpp-gguf/VoxCPM2-GGUF": {"family":"voxcpm2","files":["voxcpm2-bf16.gguf","voxcpm2-orig.gguf","voxcpm2-q8_0.gguf"],"spec":{},"task":"tts"}
+ }
+}
diff --git a/studio/backend/tests/fixtures/audio_cpp_option_golden.json b/studio/backend/tests/fixtures/audio_cpp_option_golden.json
new file mode 100644
index 00000000000..377d442bae3
--- /dev/null
+++ b/studio/backend/tests/fixtures/audio_cpp_option_golden.json
@@ -0,0 +1,189 @@
+{
+ "audio_cpp_tag": "v0.9.0-unsloth.1",
+ "about": "Expected Advanced options per catalog model and workflow, derived from audio.cpp's webui/configs/model_params.json and C++ runtime defaults independently of audio_cpp_option_table.py (which names each value's source). include: offered with this default (null: offered, sends nothing); optional: if offered, with this default; exclude: never offered. names: accepted aliases.",
+ "models": {
+ "audio-cpp/MiniMax-Music3-GGUF": [
+ {"exclude":["duration_sec","lyrics"],"include":{"ar_guidance_scale":{"default":1.5},"guidance_scale":{"default":1.7},"num_inference_steps":{"default":30},"top_k":{"default":50}},"optional":{},"workflow":"music"}
+ ],
+ "audio-cpp/Yue2-3B-GGUF": [
+ {"exclude":["export_semantic","lyrics","semantic_max_tokens","semantic_min_tokens","semantic_prefix","stop_after","style"],"include":{"abc_repetition_penalty":{"default":1.005},"abc_temperature":{"default":0.7},"abc_top_k":{"default":30},"abc_top_p":{"default":0.9},"cot":{"default":"full"},"guidance_scale":{"default":null},"num_inference_steps":{"default":8},"semantic_repetition_penalty":{"default":1.2},"semantic_temperature":{"default":1.0},"semantic_top_k":{"default":100},"semantic_top_p":{"default":0.95}},"optional":{},"workflow":"music"}
+ ],
+ "audio-cpp/audio.cpp-gguf/ACE-Step1.5-GGUF": [
+ {"exclude":["duration_seconds","guidance_scale","route"],"include":{"lm_cfg_scale":{"default":2.0},"lm_temperature":{"default":0.85},"lm_top_k":{"default":0},"lm_top_p":{"default":0.9},"num_inference_steps":{"default":8},"shift":{"default":3.0}},"optional":{"bpm":{"default":null},"keyscale":{"default":null},"timesignature":{"default":null}},"variant":"turbo","workflow":"music"},
+ {"exclude":["duration_seconds","route"],"include":{"guidance_scale":{"default":7.0},"lm_cfg_scale":{"default":2.0},"lm_temperature":{"default":0.85},"lm_top_k":{"default":0},"lm_top_p":{"default":0.9},"num_inference_steps":{"default":32},"shift":{"default":1.0}},"optional":{"bpm":{"default":null},"keyscale":{"default":null},"timesignature":{"default":null}},"variant":"base","workflow":"music"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Breeze-TTS-2-GGUF": [
+ {"exclude":["instruction"],"include":{"depth_temperature":{"default":0.9},"guidance_scale":{"default":1.0},"max_tokens":{"default":1500},"temperature":{"default":0.9},"text_chunk_mode":{"default":"default"},"text_chunk_size":{"default":600},"top_k":{"default":50},"top_p":{"default":1.0}},"optional":{},"workflow":"speak"},
+ {"exclude":["instruction","voice"],"include":{"depth_temperature":{"default":0.9},"guidance_scale":{"default":1.0},"max_tokens":{"default":1500},"temperature":{"default":0.9},"text_chunk_mode":{"default":"default"},"text_chunk_size":{"default":600},"top_k":{"default":50},"top_p":{"default":1.0}},"optional":{},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Chatterbox-GGUF": [
+ {"exclude":["voice"],"include":{"exaggeration":{"default":0.5},"guidance_scale":{"default":0.5},"min_p":{"default":0.05},"repetition_penalty":{"default":1.2},"temperature":{"default":0.8},"top_p":{"default":1.0}},"optional":{},"workflow":"clone"},
+ {"exclude":["exaggeration","min_p","repetition_penalty","temperature","top_p","voice"],"include":{"num_inference_steps":{"default":10},"s3gen_cfg_rate":{"default":0.7}},"optional":{},"workflow":"convert"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Chatterbox-Turbo-GGUF": [
+ {"exclude":["exaggeration","max_new_tokens","min_p"],"include":{"repetition_penalty":{"default":1.2},"temperature":{"default":0.8},"text_chunk_size":{"default":128},"top_k":{"default":1000},"top_p":{"default":0.95}},"optional":{},"workflow":"speak"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Confucius4-TTS-GGUF": [
+ {"exclude":["language","voice"],"include":{"cross_fade_duration_sec":{"default":0.3},"edge_fade_duration_sec":{"default":0.1},"edge_pad_duration_sec":{"default":0.1},"guidance_scale":{"default":0.7},"max_tokens":{"default":1520},"num_beams":{"default":3},"num_inference_steps":{"default":25},"repetition_penalty":{"default":10.0},"temperature":{"default":0.8},"text_chunk_mode":{"default":"default"},"text_chunk_size":{"default":80},"top_k":{"default":30},"top_p":{"default":0.8}},"optional":{},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/ControlFoley-GGUF": [
+ {"exclude":["duration_sec","mask_away_clip"],"include":{"guidance_scale":{"default":4.5},"num_inference_steps":{"default":25}},"optional":{},"workflow":"music"}
+ ],
+ "audio-cpp/audio.cpp-gguf/CosyVoice3-GGUF": [
+ {"exclude":["voice"],"include":{"max_tokens":{"default":null},"min_tokens":{"default":null},"num_inference_steps":{"default":10},"text_chunk_mode":{"default":"default"},"text_chunk_size":{"default":600},"top_k":{"default":25}},"optional":{},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/DotTTS-Edit-GGUF": [
+ {"exclude":["source_text","target_text","template_name","use_xvector"],"include":{"guidance_scale":{"default":1.2},"max_tokens":{"default":500},"num_inference_steps":{"default":10},"sampler_mode":{"default":"euler"},"speaker_scale":{"default":1.5},"text_chunk_mode":{"default":"tag_aware"},"text_chunk_size":{"default":320},"vocoder_merge_steps":{"default":4}},"optional":{"reference_duration_sec":{"default":null}},"workflow":"speak"},
+ {"exclude":["source_text","target_text","template_name","use_xvector","voice"],"include":{"guidance_scale":{"default":1.2},"max_tokens":{"default":500},"num_inference_steps":{"default":10},"sampler_mode":{"default":"euler"},"speaker_scale":{"default":1.5},"text_chunk_mode":{"default":"tag_aware"},"text_chunk_size":{"default":320},"vocoder_merge_steps":{"default":4}},"optional":{"reference_duration_sec":{"default":null}},"workflow":"clone"},
+ {"exclude":["source_text","target_text","template_name","voice"],"include":{"guidance_scale":{"default":1.2},"max_tokens":{"default":500},"num_inference_steps":{"default":10},"sampler_mode":{"default":"euler"},"speaker_scale":{"default":1.5},"text_chunk_mode":{"default":"tag_aware"},"text_chunk_size":{"default":320},"use_xvector":{"default":"auto"},"vocoder_merge_steps":{"default":4}},"optional":{"reference_duration_sec":{"default":null}},"workflow":"edit"}
+ ],
+ "audio-cpp/audio.cpp-gguf/DotTTS-MF-GGUF": [
+ {"exclude":[],"include":{"guidance_scale":{"default":1.2},"max_tokens":{"default":500},"num_inference_steps":{"default":10},"sampler_mode":{"default":"euler"},"speaker_scale":{"default":1.5},"text_chunk_mode":{"default":"tag_aware"},"text_chunk_size":{"default":320},"vocoder_merge_steps":{"default":4}},"optional":{"reference_duration_sec":{"default":null}},"workflow":"speak"},
+ {"exclude":["voice"],"include":{"guidance_scale":{"default":1.2},"max_tokens":{"default":500},"num_inference_steps":{"default":10},"sampler_mode":{"default":"euler"},"speaker_scale":{"default":1.5},"text_chunk_mode":{"default":"tag_aware"},"text_chunk_size":{"default":320},"vocoder_merge_steps":{"default":4}},"optional":{"reference_duration_sec":{"default":null}},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/DotTTS-SOAR-GGUF": [
+ {"exclude":[],"include":{"guidance_scale":{"default":1.2},"max_tokens":{"default":500},"num_inference_steps":{"default":10},"sampler_mode":{"default":"euler"},"speaker_scale":{"default":1.5},"text_chunk_mode":{"default":"tag_aware"},"text_chunk_size":{"default":320},"vocoder_merge_steps":{"default":4}},"optional":{"reference_duration_sec":{"default":null}},"workflow":"speak"},
+ {"exclude":["voice"],"include":{"guidance_scale":{"default":1.2},"max_tokens":{"default":500},"num_inference_steps":{"default":10},"sampler_mode":{"default":"euler"},"speaker_scale":{"default":1.5},"text_chunk_mode":{"default":"tag_aware"},"text_chunk_size":{"default":320},"vocoder_merge_steps":{"default":4}},"optional":{"reference_duration_sec":{"default":null}},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/DramaBox-GGUF": [
+ {"exclude":["reference_duration_sec"],"include":{"audio_chunk_duration_sec":{"default":37.0},"audio_chunk_threshold_sec":{"default":45.0},"cross_fade_duration_sec":{"default":0.05},"duration_scale":{"default":1.1},"duration_sec":{"default":0.0},"guidance_rescale":{"default":"auto"},"guidance_scale":{"default":2.5},"num_inference_steps":{"default":30},"spatio_temporal_guidance_scale":{"default":1.5}},"optional":{"negative_prompt":{"default":null}},"workflow":"speak"}
+ ],
+ "audio-cpp/audio.cpp-gguf/FireRedAudio-GGUF": [
+ {"exclude":["enable_thinking","max_new_tokens","temperature","template_name","top_k","top_p","voice"],"include":{"guidance_scale":{"default":2.0},"max_new_audio_steps":{"default":750},"max_new_text_tokens":{"default":512},"min_new_audio_steps":{"default":6},"num_inference_steps":{"default":10}},"optional":{},"workflow":"clone"},
+ {"exclude":["enable_thinking","max_new_tokens","temperature","template_name","top_k","top_p","voice"],"include":{"guidance_scale":{"default":2.0},"max_new_audio_steps":{"default":750},"max_new_text_tokens":{"default":512},"min_new_audio_steps":{"default":6},"num_inference_steps":{"default":10}},"optional":{},"workflow":"edit"}
+ ],
+ "audio-cpp/audio.cpp-gguf/FireRedTTS3-Base-GGUF": [
+ {"exclude":["template_name","voice"],"include":{"guidance_scale":{"default":2.0},"num_inference_steps":{"default":10},"stop_threshold":{"default":0.5},"text_chunk_mode":{"default":"default"},"text_chunk_size":{"default":600}},"optional":{},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/FireRedTTS3-Instruct-GGUF": [
+ {"exclude":["template_name","voice"],"include":{"guidance_scale":{"default":2.0},"num_inference_steps":{"default":10},"stop_threshold":{"default":0.5},"text_chunk_mode":{"default":"default"},"text_chunk_size":{"default":600}},"optional":{},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Fish-Audio-S2-Pro-GGUF": [
+ {"exclude":["repetition_penalty"],"include":{"max_tokens":{"default":1024},"temperature":{"default":0.8},"text_chunk_mode":{"default":"word_budget"},"text_chunk_size":{"default":200},"top_k":{"default":30},"top_p":{"default":0.8}},"optional":{},"workflow":"speak"},
+ {"exclude":["repetition_penalty","voice"],"include":{"max_tokens":{"default":1024},"temperature":{"default":0.8},"text_chunk_mode":{"default":"word_budget"},"text_chunk_size":{"default":200},"top_k":{"default":30},"top_p":{"default":0.8}},"optional":{},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/HeartMuLa-GGUF": [
+ {"exclude":["infinite_chunk_audio_duration_ms","infinite_mode","lyrics","tags","text_chunk_size"],"include":{"codec_duration_sec":{"default":29.76},"codec_guidance_scale":{"default":1.25},"guidance_scale":{"default":1.5},"num_inference_steps":{"default":10},"temperature":{"default":1.0},"top_k":{"default":50}},"optional":{},"workflow":"music"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Higgs-Audio-v3-TTS-4B-GGUF": [
+ {"exclude":["repetition_penalty"],"include":{"max_tokens":{"default":2048},"temperature":{"default":0.8},"text_chunk_size":{"default":1024},"top_k":{"default":30},"top_p":{"default":0.8}},"optional":{},"workflow":"speak"},
+ {"exclude":["repetition_penalty","voice"],"include":{"max_tokens":{"default":2048},"temperature":{"default":0.8},"text_chunk_size":{"default":1024},"top_k":{"default":30},"top_p":{"default":0.8}},"optional":{},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/IndexTTS2-GGUF": [
+ {"exclude":["voice"],"include":{"do_sample":{"default":true},"duration_factor":{"default":1.0},"emotion_alpha":{"default":1.0},"interval_silence_ms":{"default":200},"length_penalty":{"default":0.0},"max_tokens":{"default":1500},"num_beams":{"default":3},"repetition_penalty":{"default":10.0},"temperature":{"default":0.8},"top_k":{"default":30},"top_p":{"default":0.8},"use_random_emotion":{"default":false}},"optional":{},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/IndexTTS2.5-GGUF": [
+ {"exclude":["voice"],"include":{"do_sample":{"default":true},"duration_factor":{"default":1.0},"emotion_alpha":{"default":1.0},"interval_silence_ms":{"default":200},"length_penalty":{"default":0.0},"max_tokens":{"default":1500},"num_beams":{"default":3},"repetition_penalty":{"default":10.0},"temperature":{"default":0.8},"top_k":{"default":30},"top_p":{"default":0.8},"use_random_emotion":{"default":false}},"optional":{},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Inflect-Micro-v2-GGUF": [
+ {"exclude":["text_chunk_mode"],"include":{"speaking_rate":{"default":1.0},"text_chunk_size":{"default":280},"variation":{"default":0.667}},"optional":{},"workflow":"speak"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Irodori-TTS-500M-v3-GGUF": [
+ {"exclude":["caption","caption_guidance_scale","no_ref"],"include":{"duration_scale":{"default":1.0},"duration_sec":{"default":null},"guidance_max_t":{"default":1.0},"guidance_min_t":{"default":0.5},"guidance_mode":{"default":"independent"},"guidance_scale":{"default":null},"max_duration_sec":{"default":30.0},"min_duration_sec":{"default":0.5},"num_inference_steps":{"default":40},"speaker_guidance_scale":{"default":5.0},"text_chunk_mode":{"default":"endline"},"text_chunk_size":{"default":null},"text_guidance_scale":{"default":3.0},"trim_tail":{"default":true}},"optional":{},"workflow":"speak"},
+ {"exclude":["caption","caption_guidance_scale","no_ref","voice"],"include":{"duration_scale":{"default":1.0},"duration_sec":{"default":null},"guidance_max_t":{"default":1.0},"guidance_min_t":{"default":0.5},"guidance_mode":{"default":"independent"},"guidance_scale":{"default":null},"max_duration_sec":{"default":30.0},"min_duration_sec":{"default":0.5},"num_inference_steps":{"default":40},"speaker_guidance_scale":{"default":5.0},"text_chunk_mode":{"default":"endline"},"text_chunk_size":{"default":null},"text_guidance_scale":{"default":3.0},"trim_tail":{"default":true}},"optional":{},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Irodori-TTS-600M-v3-VoiceDesign-GGUF": [
+ {"exclude":["no_ref"],"include":{"caption_guidance_scale":{"default":3.0},"duration_scale":{"default":1.0},"duration_sec":{"default":null},"guidance_max_t":{"default":1.0},"guidance_min_t":{"default":0.5},"guidance_mode":{"default":"independent"},"guidance_scale":{"default":null},"max_duration_sec":{"default":30.0},"min_duration_sec":{"default":0.5},"num_inference_steps":{"default":40},"speaker_guidance_scale":{"default":5.0},"text_chunk_mode":{"default":"endline"},"text_chunk_size":{"default":null},"text_guidance_scale":{"default":3.0},"trim_tail":{"default":true}},"optional":{},"workflow":"speak"},
+ {"exclude":["no_ref","voice"],"include":{"caption_guidance_scale":{"default":3.0},"duration_scale":{"default":1.0},"duration_sec":{"default":null},"guidance_max_t":{"default":1.0},"guidance_min_t":{"default":0.5},"guidance_mode":{"default":"independent"},"guidance_scale":{"default":null},"max_duration_sec":{"default":30.0},"min_duration_sec":{"default":0.5},"num_inference_steps":{"default":40},"speaker_guidance_scale":{"default":5.0},"text_chunk_mode":{"default":"endline"},"text_chunk_size":{"default":null},"text_guidance_scale":{"default":3.0},"trim_tail":{"default":true}},"optional":{},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Irodori-TTS-v4-Small-GGUF": [
+ {"exclude":["no_ref"],"include":{"caption_guidance_scale":{"default":3.0},"duration_scale":{"default":1.0},"duration_sec":{"default":null},"guidance_max_t":{"default":1.0},"guidance_min_t":{"default":0.5},"guidance_mode":{"default":"independent"},"guidance_scale":{"default":null},"max_duration_sec":{"default":30.0},"min_duration_sec":{"default":0.5},"num_inference_steps":{"default":40},"speaker_guidance_scale":{"default":5.0},"text_chunk_mode":{"default":"endline"},"text_chunk_size":{"default":null},"text_guidance_scale":{"default":3.0},"trim_tail":{"default":true}},"optional":{},"workflow":"speak"},
+ {"exclude":["no_ref","voice"],"include":{"caption_guidance_scale":{"default":3.0},"duration_scale":{"default":1.0},"duration_sec":{"default":null},"guidance_max_t":{"default":1.0},"guidance_min_t":{"default":0.5},"guidance_mode":{"default":"independent"},"guidance_scale":{"default":null},"max_duration_sec":{"default":30.0},"min_duration_sec":{"default":0.5},"num_inference_steps":{"default":40},"speaker_guidance_scale":{"default":5.0},"text_chunk_mode":{"default":"endline"},"text_chunk_size":{"default":null},"text_guidance_scale":{"default":3.0},"trim_tail":{"default":true}},"optional":{},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/KittenTTS-GGUF": [
+ {"exclude":[],"include":{"speed":{"default":1.0},"text_chunk_mode":{"default":"default"},"text_chunk_size":{"default":400},"voice":{"default":"Leo"}},"optional":{},"workflow":"speak"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Kokoro-82M-GGUF": [
+ {"exclude":["language","return_timestamps","seed"],"include":{"speed":{"default":1.0},"text_chunk_size":{"default":240},"voice":{"default":"af_heart"}},"optional":{},"workflow":"speak"}
+ ],
+ "audio-cpp/audio.cpp-gguf/MOSS-TTS-Local-v1.5-GGUF": [
+ {"exclude":[],"include":{"do_sample":{"default":true},"max_tokens":{"default":4096},"repetition_penalty":{"default":1.0},"temperature":{"default":1.7},"text_temperature":{"default":1.0},"top_k":{"default":25},"top_p":{"default":0.8}},"optional":{},"workflow":"speak"},
+ {"exclude":["voice"],"include":{"do_sample":{"default":true},"max_tokens":{"default":4096},"repetition_penalty":{"default":1.0},"temperature":{"default":1.7},"text_temperature":{"default":1.0},"top_k":{"default":25},"top_p":{"default":0.8}},"optional":{},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/MOSS-TTS-Nano-100M-GGUF": [
+ {"exclude":[],"include":{"do_sample":{"default":true},"repetition_penalty":{"default":1.0},"temperature":{"default":1.7},"text_temperature":{"default":1.5},"top_k":{"default":25},"top_p":{"default":0.8}},"optional":{"max_tokens":{"default":null}},"workflow":"speak"},
+ {"exclude":["voice"],"include":{"do_sample":{"default":true},"repetition_penalty":{"default":1.0},"temperature":{"default":1.7},"text_temperature":{"default":1.5},"top_k":{"default":25},"top_p":{"default":0.8}},"optional":{"max_tokens":{"default":null}},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/MOSS-VoiceGenerator-GGUF": [
+ {"exclude":["voice"],"include":{"repetition_penalty":{"default":1.1},"temperature":{"default":1.5},"top_k":{"default":50},"top_p":{"default":0.6}},"optional":{"max_frames":{"default":null},"min_frames":{"default":null}},"workflow":"speak"}
+ ],
+ "audio-cpp/audio.cpp-gguf/MagpieTTS-Multilingual-357M-GGUF": [
+ {"exclude":[],"include":{"guidance_scale":{"default":2.5},"max_tokens":{"default":500},"temperature":{"default":0.6},"text_chunk_mode":{"default":"default"},"text_chunk_size":{"default":300},"top_k":{"default":80}},"optional":{},"workflow":"speak"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Maya1-GGUF": [
+ {"exclude":["instruct"],"include":{"max_tokens":{"default":2048},"min_tokens":{"default":28},"repetition_penalty":{"default":1.1},"temperature":{"default":0.4},"text_chunk_mode":{"default":"tag_aware"},"text_chunk_size":{"default":300},"top_p":{"default":0.9}},"optional":{},"workflow":"speak"}
+ ],
+ "audio-cpp/audio.cpp-gguf/MeanVC2-GGUF": [
+ {"exclude":["seed","voice"],"include":{},"optional":{},"workflow":"convert"}
+ ],
+ "audio-cpp/audio.cpp-gguf/MiDashengLM-Gen-GGUF": [
+ {"exclude":["duration_sec"],"include":{"guidance_scale":{"default":2.0},"min_stop_step":{"default":5},"stop_threshold":{"default":0.5}},"optional":{},"workflow":"music"}
+ ],
+ "audio-cpp/audio.cpp-gguf/MioTTS-1.7B-GGUF": [
+ {"exclude":["best_of_n","voice"],"include":{"do_sample":{"default":true},"frequency_penalty":{"default":0.0},"max_tokens":{"default":700},"presence_penalty":{"default":0.0},"repetition_penalty":{"default":1.0},"temperature":{"default":0.8},"top_k":{"default":50},"top_p":{"default":1.0}},"optional":{},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/NeuTTS-2E-GGUF": [
+ {"exclude":[],"include":{"emotion":{"default":"neutral"},"max_tokens":{"default":0},"min_tokens":{"default":50},"temperature":{"default":1.0},"text_chunk_mode":{"default":"default"},"text_chunk_size":{"default":600},"top_k":{"default":50},"voice_id":{"default":"emily"}},"optional":{},"workflow":"speak"}
+ ],
+ "audio-cpp/audio.cpp-gguf/OmniVoice-GGUF": [
+ {"exclude":[],"include":{"audio_chunk_duration":{"default":15.0},"audio_chunk_threshold":{"default":30.0},"class_temperature":{"default":0.0},"denoise":{"default":true},"guidance_scale":{"default":2.0},"layer_penalty_factor":{"default":5.0},"num_inference_steps":{"default":32},"position_temperature":{"default":5.0},"postprocess_output":{"default":true},"preprocess_prompt":{"default":true},"speed":{"default":1.0},"t_shift":{"default":0.1}},"optional":{"duration":{"default":null}},"workflow":"speak"},
+ {"exclude":["voice"],"include":{"audio_chunk_duration":{"default":15.0},"audio_chunk_threshold":{"default":30.0},"class_temperature":{"default":0.0},"denoise":{"default":true},"guidance_scale":{"default":2.0},"layer_penalty_factor":{"default":5.0},"num_inference_steps":{"default":32},"position_temperature":{"default":5.0},"postprocess_output":{"default":true},"preprocess_prompt":{"default":true},"speed":{"default":1.0},"t_shift":{"default":0.1}},"optional":{"duration":{"default":null}},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Piper-TTS-GGUF": [
+ {"exclude":["seed"],"include":{"duration_variation":{"default":0.8},"speed":{"default":1.0},"text_chunk_mode":{"default":"default"},"text_chunk_size":{"default":280},"variation":{"default":0.667}},"optional":{},"workflow":"speak"}
+ ],
+ "audio-cpp/audio.cpp-gguf/PocketTTS-GGUF": [
+ {"exclude":[],"include":{"eos_threshold":{"default":-4.0},"frames_after_eos":{"default":-1}},"optional":{"noise_clamp":{"default":null},"temperature":{"default":null},"voice":{"default":"alba"}},"workflow":"speak"},
+ {"exclude":["voice"],"include":{"eos_threshold":{"default":-4.0},"frames_after_eos":{"default":-1}},"optional":{"noise_clamp":{"default":null},"temperature":{"default":null}},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Qwen3-TTS-12Hz-0.6B-Base-GGUF": [
+ {"exclude":["voice"],"include":{"do_sample":{"default":true},"repetition_penalty":{"default":1.05},"subtalker_do_sample":{"default":true},"subtalker_temperature":{"default":0.9},"subtalker_top_k":{"default":50},"subtalker_top_p":{"default":1.0},"temperature":{"default":0.9},"text_chunk_size":{"default":8192},"top_k":{"default":50},"top_p":{"default":1.0},"x_vector_only_mode":{"default":false}},"optional":{"max_tokens":{"default":null}},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Qwen3-TTS-12Hz-1.7B-Base-GGUF": [
+ {"exclude":["voice"],"include":{"do_sample":{"default":true},"repetition_penalty":{"default":1.05},"subtalker_do_sample":{"default":true},"subtalker_temperature":{"default":0.9},"subtalker_top_k":{"default":50},"subtalker_top_p":{"default":1.0},"temperature":{"default":0.9},"text_chunk_size":{"default":8192},"top_k":{"default":50},"top_p":{"default":1.0},"x_vector_only_mode":{"default":false}},"optional":{"max_tokens":{"default":null}},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Qwen3-TTS-12Hz-1.7B-CustomVoice-GGUF": [
+ {"exclude":[],"include":{"do_sample":{"default":true},"repetition_penalty":{"default":1.05},"subtalker_do_sample":{"default":true},"subtalker_temperature":{"default":0.9},"subtalker_top_k":{"default":50},"subtalker_top_p":{"default":1.0},"temperature":{"default":0.9},"text_chunk_size":{"default":8192},"top_k":{"default":50},"top_p":{"default":1.0},"voice":{"default":"vivian"}},"optional":{"max_tokens":{"default":null}},"workflow":"speak"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Qwen3-TTS-12Hz-1.7B-VoiceDesign-GGUF": [
+ {"exclude":["voice"],"include":{"do_sample":{"default":true},"repetition_penalty":{"default":1.05},"subtalker_do_sample":{"default":true},"subtalker_temperature":{"default":0.9},"subtalker_top_k":{"default":50},"subtalker_top_p":{"default":1.0},"temperature":{"default":0.9},"text_chunk_size":{"default":8192},"top_k":{"default":50},"top_p":{"default":1.0}},"optional":{"max_tokens":{"default":null}},"workflow":"speak"}
+ ],
+ "audio-cpp/audio.cpp-gguf/RVC-GGUF": [
+ {"exclude":["audio_pad_duration_sec","pitch_extractor","semitone_shift","speaker_id","voice","voice_id"],"include":{"output_sample_rate":{"default":0},"pitch_filter_radius":{"default":3},"retrieval_blend":{"default":0.0},"rms_mix_rate":{"default":0.25},"split_center_sec":{"default":30},"split_query_sec":{"default":5},"split_threshold_sec":{"default":32},"unvoiced_protection":{"default":0.33}},"optional":{},"workflow":"convert"}
+ ],
+ "audio-cpp/audio.cpp-gguf/SeedVC-MLX-GGUF": [
+ {"exclude":["auto_f0_adjust","f0_condition","semitone_shift","voice"],"include":{"inference_guidance_scale":{"default":0.7},"intelligibility_guidance_scale":{"default":0.7},"length_adjust":{"default":1.0},"num_inference_steps":{"default":30},"similarity_guidance_scale":{"default":0.7},"voice_anonymization":{"default":false}},"optional":{"route":{"default":null}},"workflow":"convert"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Stable-Audio-3-Medium-GGUF": [
+ {"exclude":["batch_size","duration_seconds"],"include":{"apg_scale":{"default":1.0},"guidance_scale":{"default":1.0},"num_inference_steps":{"default":8}},"optional":{"negative_prompt":{"default":null}},"workflow":"music"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Stable-Audio-3-Small-Music-GGUF": [
+ {"exclude":["batch_size","duration_seconds"],"include":{"apg_scale":{"default":1.0},"guidance_scale":{"default":1.0},"num_inference_steps":{"default":8}},"optional":{"negative_prompt":{"default":null}},"workflow":"music"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Stable-Audio-3-Small-SFX-GGUF": [
+ {"exclude":["batch_size","duration_seconds"],"include":{"apg_scale":{"default":1.0},"guidance_scale":{"default":1.0},"num_inference_steps":{"default":8}},"optional":{"negative_prompt":{"default":null}},"workflow":"music"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Supertonic-3-GGUF": [
+ {"exclude":[],"include":{"num_inference_steps":{"default":8},"speed":{"default":1.05,"names":["speed","speaking_rate"]},"voice":{"default":"M1","values":["M1","M2","M3","M4","M5","F1","F2","F3","F4","F5"]}},"optional":{"text_chunk_size":{"default":null}},"workflow":"speak"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Tone-Color-VC-GGUF": [
+ {"exclude":["voice"],"include":{"temperature":{"default":0.3}},"optional":{},"workflow":"convert"}
+ ],
+ "audio-cpp/audio.cpp-gguf/Vevo2-GGUF": [
+ {"exclude":["audio_chunk_duration_sec","cross_fade_duration_sec","predict_target_prosody","reference_duration_seconds","route","target_duration_seconds","use_prosody_code","voice"],"include":{"max_tokens":{"default":500},"repetition_penalty":{"default":1.1},"temperature":{"default":0.7},"top_k":{"default":20},"top_p":{"default":0.8}},"optional":{},"workflow":"clone"},
+ {"exclude":["predict_target_prosody","reference_duration_seconds","route","target_duration_seconds","use_prosody_code","voice"],"include":{"audio_chunk_duration_sec":{"default":0.0},"cross_fade_duration_sec":{"default":1.0},"max_tokens":{"default":500},"repetition_penalty":{"default":1.1},"temperature":{"default":0.7},"top_k":{"default":20},"top_p":{"default":0.8}},"optional":{},"workflow":"edit"},
+ {"exclude":["predict_target_prosody","reference_duration_seconds","route","target_duration_seconds","use_pitch_shift","use_prosody_code","voice"],"include":{"num_inference_steps":{"default":32}},"optional":{},"workflow":"convert"}
+ ],
+ "audio-cpp/audio.cpp-gguf/VibeVoice-1.5B-GGUF": [
+ {"exclude":[],"include":{"do_sample":{"default":false},"guidance_scale":{"default":1.3},"max_length_times":{"default":2.0},"num_inference_steps":{"default":10},"temperature":{"default":1.0},"top_k":{"default":50},"top_p":{"default":1.0}},"optional":{},"workflow":"speak"}
+ ],
+ "audio-cpp/audio.cpp-gguf/VoxCPM1-GGUF": [
+ {"exclude":["prompt_text"],"include":{"guidance_scale":{"default":2.0},"max_tokens":{"default":4096},"min_tokens":{"default":2},"num_inference_steps":{"default":10},"retry_badcase":{"default":true},"retry_badcase_max_times":{"default":3},"retry_badcase_ratio_threshold":{"default":6.0},"text_chunk_mode":{"default":"tag_aware"}},"optional":{},"workflow":"speak"},
+ {"exclude":["prompt_text","voice"],"include":{"guidance_scale":{"default":2.0},"max_tokens":{"default":4096},"min_tokens":{"default":2},"num_inference_steps":{"default":10},"retry_badcase":{"default":true},"retry_badcase_max_times":{"default":3},"retry_badcase_ratio_threshold":{"default":6.0},"text_chunk_mode":{"default":"tag_aware"}},"optional":{},"workflow":"clone"}
+ ],
+ "audio-cpp/audio.cpp-gguf/VoxCPM2-GGUF": [
+ {"exclude":["prompt_text"],"include":{"guidance_scale":{"default":2.0},"min_tokens":{"default":2},"num_inference_steps":{"default":10},"retry_badcase":{"default":true},"retry_badcase_max_times":{"default":3},"retry_badcase_ratio_threshold":{"default":6.0},"text_chunk_mode":{"default":"tag_aware"}},"optional":{"max_tokens":{"default":null}},"workflow":"speak"},
+ {"exclude":["prompt_text","voice"],"include":{"guidance_scale":{"default":2.0},"min_tokens":{"default":2},"num_inference_steps":{"default":10},"retry_badcase":{"default":true},"retry_badcase_max_times":{"default":3},"retry_badcase_ratio_threshold":{"default":6.0},"text_chunk_mode":{"default":"tag_aware"}},"optional":{"max_tokens":{"default":null}},"workflow":"clone"}
+ ]
+ }
+}
diff --git a/studio/backend/tests/test_audio_cpp_convert_requests.py b/studio/backend/tests/test_audio_cpp_convert_requests.py
index 4bbf20cc277..2ec4d457b01 100644
--- a/studio/backend/tests/test_audio_cpp_convert_requests.py
+++ b/studio/backend/tests/test_audio_cpp_convert_requests.py
@@ -166,6 +166,10 @@ def _convert(
return body["request"]
+# Recommended RVC defaults Unsloth Studio fills when the user leaves them unset.
+_RVC_DEFAULTS = {"retrieval_blend": "0", "rms_mix_rate": "0.25", "unvoiced_protection": "0.33"}
+
+
def test_rvc_converts_to_a_builtin_voice_by_options_voice_id(started):
backend = _backend(_model("rvc", _RVC_OPTIONS), started)
request = _convert(
@@ -178,15 +182,21 @@ def test_rvc_converts_to_a_builtin_voice_by_options_voice_id(started):
# Strict: no seed, no voice_ref, no language; the typed fields win over driven options.
assert request == {
"audio": SOURCE,
- "options": {"retrieval_blend": "0.6", "voice_id": "manthos", "semitone_shift": "3"},
+ "options": {
+ "retrieval_blend": "0.6",
+ "voice_id": "manthos",
+ "semitone_shift": "3",
+ "rms_mix_rate": "0.25",
+ "unvoiced_protection": "0.33",
+ },
}
assert _convert(backend, inputs = {"source": SOURCE}) == {
"audio": SOURCE,
- "options": {"voice_id": "default"},
+ "options": {"voice_id": "default", **_RVC_DEFAULTS},
}
# Auto is not offered on RVC, so it never becomes auto_f0_adjust.
request = _convert(backend, inputs = {"source": SOURCE}, pitch = -5, pitch_auto = True)
- assert request["options"] == {"voice_id": "default", "semitone_shift": "-5"}
+ assert request["options"] == {"voice_id": "default", "semitone_shift": "-5", **_RVC_DEFAULTS}
assert started == []
@@ -207,6 +217,14 @@ def test_a_reference_family_sends_audio_voice_ref_and_seed(family):
assert request == {"audio": SOURCE, "voice_ref": TARGET, "seed": "7"}
+# Recommended Seed-VC defaults Unsloth Studio fills when the user leaves them unset.
+_SEED_VC_DEFAULTS = {
+ "length_adjust": "1",
+ "similarity_guidance_scale": "0.7",
+ "voice_anonymization": "false",
+}
+
+
def test_seed_vc_speech_sends_no_pitch_and_its_route_by_model_entry(started):
model = _model("seed_vc", _SEED_VC_OPTIONS)
backend = _backend(model, started)
@@ -225,7 +243,11 @@ def test_seed_vc_speech_sends_no_pitch_and_its_route_by_model_entry(started):
assert request == {
"audio": SOURCE,
"voice_ref": TARGET,
- "options": {"length_adjust": "1.25", "similarity_guidance_scale": "1.2"},
+ "options": {
+ "length_adjust": "1.25",
+ "similarity_guidance_scale": "1.2",
+ "voice_anonymization": "false",
+ },
"seed": "7",
}
# The default engine (v2_vc) runs on the loaded session: no restart, no route sent.
@@ -249,16 +271,20 @@ def test_seed_vc_singing_reloads_under_svc_and_sends_its_pitch(started):
backend = _backend(model, started)
# The runtime applies the manual shift on top of the matched pitch, as seed-vc does.
request = _convert(backend, mode = "singing", pitch = 12, pitch_auto = True)
- assert request["options"] == {"auto_f0_adjust": "true", "semitone_shift": "12"}
+ assert request["options"] == {
+ "auto_f0_adjust": "true",
+ "semitone_shift": "12",
+ **_SEED_VC_DEFAULTS,
+ }
request = _convert(backend, mode = "singing", pitch_auto = True)
- assert request["options"] == {"auto_f0_adjust": "true"}
+ assert request["options"] == {"auto_f0_adjust": "true", **_SEED_VC_DEFAULTS}
(server,) = started
assert server.model.server_task == "svc"
assert backend.models[model.id]["audio_server_task"] == "svc"
assert backend.models[model.id]["audio_convert_route"] == "v1_svc"
# A speech route sent in singing is dropped: v1_svc is the only singing route.
request = _convert(backend, mode = "singing", pitch = -3, options = {"route": "v1_xlsr_hift_vc"})
- assert request["options"] == {"semitone_shift": "-3"}
+ assert request["options"] == {"semitone_shift": "-3", **_SEED_VC_DEFAULTS}
assert len(started) == 1
_convert(backend)
assert [s.model.server_task for s in started] == ["svc", "vc"]
diff --git a/studio/backend/tests/test_audio_cpp_models.py b/studio/backend/tests/test_audio_cpp_models.py
index 83d18a4109e..8057a74b43c 100644
--- a/studio/backend/tests/test_audio_cpp_models.py
+++ b/studio/backend/tests/test_audio_cpp_models.py
@@ -1542,7 +1542,11 @@ def test_audio_options_reach_the_request_validated():
"hi", audio_options = {"voice": "am_adam", "speed": 1.5, "bogus": 1, "mode": "zzz"}
)
body = srv.calls[0][1]
- assert body["voice"] == "am_adam" and body["options"] == {"speed": 1.5}
+ assert body["voice"] == "am_adam" and body["options"] == {
+ "speed": 1.5,
+ "mode": "a",
+ "text_chunk_size": 240,
+ }
def test_speech_retry_keeps_package_defaults_and_user_options_and_covers_500():
@@ -1607,7 +1611,13 @@ def test_minimax_takes_the_caption_as_input_and_lyrics_and_duration_as_options()
request = srv.calls[0][1]["request"]
assert request["text"] == "lofi" and request["seed"] == "3"
# The runtime rejects duration_sec beside duration_seconds ("conflicting option values").
- assert request["options"] == {"top_k": 9, "lyrics": "[verse] hi"}
+ assert request["options"] == {
+ "top_k": 9,
+ "lyrics": "[verse] hi",
+ "ar_guidance_scale": 1.5,
+ "guidance_scale": 1.7,
+ "num_inference_steps": 30,
+ }
assert request["duration_seconds"] == 20.0
with pytest.raises(RuntimeError, match = "needs lyrics"):
b.generate_audio_response("", instructions = "lofi")
diff --git a/studio/backend/tests/test_audio_cpp_music_requests.py b/studio/backend/tests/test_audio_cpp_music_requests.py
index 22ec0637b62..4668724f68d 100644
--- a/studio/backend/tests/test_audio_cpp_music_requests.py
+++ b/studio/backend/tests/test_audio_cpp_music_requests.py
@@ -290,7 +290,7 @@ def test_status_reports_the_music_studio_rules():
"max_ranges": 1,
"max_source_s": 240.0,
}
- # Upstream ACE-Step 1.5 offers "complete" and 1-200 steps on base only; turbo keeps 1-20.
+ # Upstream ACE-Step 1.5 offers "complete" and 1-200 steps on base only; turbo keeps 1-8.
base = _ace_base()
assert audio_cpp_backend.model_info_fields(base)["audio_music"]["modes"][1]["actions"] == [
"repaint",
@@ -300,9 +300,11 @@ def test_status_reports_the_music_studio_rules():
]
steps = {o["name"]: o for o in base.options}["num_inference_steps"]
assert (steps["max"], steps["default"]) == (200, 32)
- turbo = base.with_variant(AudioCppVariant("turbo/Q8_0", (), "ACE-Step1.5-GGUF/turbo/t.gguf"))
+ turbo = base.with_variant(
+ AudioCppVariant("turbo/Q8_0", (), "ACE-Step1.5-GGUF/turbo/ace-step-1.5-turbo-q8_0.gguf")
+ )
assert "continue" not in turbo.music.mode("edit").actions
- assert {o["name"]: o for o in turbo.options}["num_inference_steps"]["max"] == 20
+ assert {o["name"]: o for o in turbo.options}["num_inference_steps"]["max"] == 8
stable = cm.music_rules(_model("stable_audio", "Stable-Audio-3-Medium-GGUF", strict = False))
assert [m["id"] for m in stable["modes"]] == ["song", "sfx", "edit"]
assert stable["modes"][0]["variations"] == {"max": 4, "how": "batch", "loaded": 1}
@@ -395,7 +397,14 @@ def test_minimax_gguf_keeps_its_body_through_the_studio():
"text": "lofi",
"lyrics": "[verse] hi",
"duration_seconds": 20.0,
- "options": {"lyrics": "[verse] hi"},
+ # Recommended MiniMax defaults ride along with Unsloth Studio's own lyrics option.
+ "options": {
+ "lyrics": "[verse] hi",
+ "num_inference_steps": 30,
+ "guidance_scale": 1.7,
+ "ar_guidance_scale": 1.5,
+ "top_k": 50,
+ },
"seed": "3",
}
assert "duration_sec" not in set(_keys(request))
@@ -414,7 +423,7 @@ def test_ace_step_song_and_instrumental():
"text": "synth pop",
"duration_seconds": 45.0,
"lyrics": "[verse] la",
- "options": {"bpm": 100, "timesignature": "3"},
+ "options": {"bpm": 100, "timesignature": "3", "num_inference_steps": 8},
"seed": "7",
}
_run(b, text = "synth pop", lyrics = "[verse] la", instrumental = True)
@@ -435,7 +444,14 @@ def test_stable_audio_batches_variations_as_a_string(starts, tmp_path):
run.mkdir()
wav, rate = _run(b, text = "house", duration_s = 6, variations = 3, seed = None, output_dir = str(run))
request = b._server.calls[0][1]["request"]
- assert request["options"] == {"batch_size": "3"} and "seed" not in request
+ assert (
+ request["options"]
+ == {
+ "batch_size": "3",
+ **STABLE_AUDIO_DEFAULTS,
+ }
+ and "seed" not in request
+ )
assert rate == 44100 and wav[:4] == b"RIFF"
manifest = json.loads((run / "outputs.json").read_text())
assert [m["id"] for m in manifest] == ["audio_0"] # the fake answers one take
@@ -553,15 +569,29 @@ def test_ace_step_repaint_has_no_duration_and_may_pass_the_end(source):
"repainting_start": 5.0,
"repainting_end": 14.0,
"repaint_strength": 0.4,
+ "num_inference_steps": 8,
},
"seed": "7",
}
assert "duration_seconds" not in set(_keys(request))
+# Recommended defaults Unsloth Studio fills when the user leaves them unset.
+STABLE_AUDIO_DEFAULTS = {"num_inference_steps": 8, "guidance_scale": 1.0, "apg_scale": 1.0}
+ACE_BASE_DEFAULTS = {
+ "num_inference_steps": 32,
+ "guidance_scale": 7.0,
+ "shift": 1.0,
+ "lm_temperature": 0.85,
+ "lm_cfg_scale": 2.0,
+ "lm_top_p": 0.9,
+ "lm_top_k": 0,
+}
+
+
def _ace_base():
return _model("ace_step", strict = False).with_variant(
- AudioCppVariant("base/Q8_0", (), "ACE-Step1.5-GGUF/base/b.gguf")
+ AudioCppVariant("base/Q8_0", (), "ACE-Step1.5-GGUF/base/ace-step-1.5-base-q8_0.gguf")
)
@@ -572,14 +602,23 @@ def test_ace_step_extend_cover_and_continue(source):
"route": "repaint",
"repainting_start": 10.0,
"repainting_end": 22.0,
+ **ACE_BASE_DEFAULTS,
}
# strength is how much to change, so a light cover keeps most source-conditioned steps.
_run(b, mode = "edit", text = "jazz", source = source, edit = {"action": "cover", "strength": 0.25})
- assert _request(b, 1)["options"] == {"route": "cover", "audio_cover_strength": 0.75}
+ assert _request(b, 1)["options"] == {
+ "route": "cover",
+ "audio_cover_strength": 0.75,
+ **ACE_BASE_DEFAULTS,
+ }
_run(
b, mode = "edit", text = "add drums", source = source, duration_s = 30, edit = {"action": "continue"}
)
- assert _request(b, 2)["options"] == {"route": "complete", "duration_seconds": 30.0}
+ assert _request(b, 2)["options"] == {
+ "route": "complete",
+ "duration_seconds": 30.0,
+ **ACE_BASE_DEFAULTS,
+ }
def test_stable_audio_inpaint_sorts_and_merges_ranges(source):
@@ -602,6 +641,7 @@ def test_stable_audio_inpaint_sorts_and_merges_ranges(source):
"audio_input_kind": "inpaint_audio",
"inpaint_mask_start_seconds": "1,5,9.5",
"inpaint_mask_end_seconds": "2,6.5,10",
+ **STABLE_AUDIO_DEFAULTS,
}
assert request["duration_seconds"] == 10.0 and request["audio"] == source
@@ -609,7 +649,11 @@ def test_stable_audio_inpaint_sorts_and_merges_ranges(source):
def test_stable_audio_restyle_sends_init_noise(source):
b = _backend(_model("stable_audio", strict = False))
_run(b, mode = "edit", text = "lofi", source = source, edit = {"action": "restyle", "strength": 0.45})
- assert _request(b)["options"] == {"audio_input_kind": "init_audio", "init_noise_level": 0.45}
+ assert _request(b)["options"] == {
+ "audio_input_kind": "init_audio",
+ "init_noise_level": 0.45,
+ **STABLE_AUDIO_DEFAULTS,
+ }
def test_an_edit_the_family_does_not_offer_fails_before_any_call(source):
diff --git a/studio/backend/tests/test_audio_cpp_option_golden.py b/studio/backend/tests/test_audio_cpp_option_golden.py
new file mode 100644
index 00000000000..c60d12081ef
--- /dev/null
+++ b/studio/backend/tests/test_audio_cpp_option_golden.py
@@ -0,0 +1,127 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Each catalog model's Advanced options against a golden written separately from the option table.
+
+``fixtures/audio_cpp_option_golden.json`` was derived from the audio.cpp webui parameter table and
+the C++ runtime defaults, without reading ``audio_cpp_option_table.py``, so a mistake has to be made
+twice to slip through. Per model and workflow it names the options that must be offered (and the
+value each starts from), the ones that may be offered (and then with which value), and the ones
+that must not be. A case id is ``<folder>[/variant]::<workflow>``."""
+
+import json
+import sys
+from pathlib import Path
+
+import pytest
+
+from core.inference import audio_cpp_models as acm
+
+_TESTS_DIR = str(Path(__file__).resolve().parent)
+if _TESTS_DIR not in sys.path:
+ sys.path.insert(0, _TESTS_DIR)
+
+from test_audio_cpp_option_table import SPECS, _resolved, catalog_hub # noqa: E402, F401
+
+GOLDEN = json.loads(
+ (Path(__file__).parent / "fixtures" / "audio_cpp_option_golden.json").read_text(
+ encoding = "utf-8"
+ )
+)["models"]
+
+
+def _cases():
+ out = []
+ for model_id, cases in sorted(GOLDEN.items()):
+ folder = model_id.rsplit("/", 1)[-1]
+ for case in cases:
+ label = folder + (f"/{case['variant']}" if case.get("variant") else "")
+ out.append(pytest.param(model_id, case, id = f"{label}::{case['workflow']}"))
+ return out
+
+
+def _variants(model_id, case):
+ models = _resolved(model_id)
+ prefix = case.get("variant")
+ if prefix:
+ models = [m for m in models if m.variant.key.lower().startswith(prefix.lower() + "/")]
+ assert models, (model_id, prefix)
+ return models
+
+
+def _same(got, want):
+ if isinstance(want, bool) or isinstance(got, bool):
+ return type(got) is type(want) and got == want
+ if isinstance(want, (int, float)) and isinstance(got, (int, float)):
+ return abs(float(got) - float(want)) <= 1e-9
+ return got == want
+
+
+def _pick(by_name, name, spec):
+ """The offered option for a golden entry (one of its accepted names), or None."""
+ names = [n for n in spec.get("names", [name]) if n in by_name]
+ assert len(names) <= 1, f"{name}: more than one alias offered {names}"
+ return by_name[names[0]] if names else None
+
+
[email protected]("model_id, case", _cases())
+def test_options_match_the_golden(catalog_hub, model_id, case):
+ workflow = case["workflow"]
+ for model in _variants(model_id, case):
+ where = f"{model_id} [{model.variant.key}] {workflow}"
+ assert workflow in model.workflows, f"{where}: workflow not offered"
+ schema = acm.workflow_options(model, workflow)
+ by_name = {o["name"]: o for o in schema}
+ sent = acm.filled_options(model, workflow, {})
+ # A tool panel sends its own options; the request fills only the model's.
+ own = {o["name"] for o in model.options}
+ problems = []
+ checked = [(n, s, True) for n, s in case["include"].items()]
+ checked += [(n, s, False) for n, s in case["optional"].items()]
+ for name, spec, required in checked:
+ option = _pick(by_name, name, spec)
+ if option is None:
+ if required:
+ problems.append(f"{name}: not offered (want default {spec['default']!r})")
+ continue
+ want = spec["default"]
+ if not _same(option.get("default"), want):
+ problems.append(f"{name}: default {option.get('default')!r}, want {want!r}")
+ if "values" in spec and sorted(option.get("values") or []) != sorted(spec["values"]):
+ problems.append(f"{name}: values {option.get('values')!r}, want {spec['values']!r}")
+ real = option["name"]
+ if real not in own:
+ if real in sent:
+ problems.append(f"{real}: a tool panel's option was filled")
+ continue
+ if want is None or real == "voice":
+ if real in sent:
+ problems.append(f"{real}: sent {sent[real]!r} though nothing should be")
+ elif model.request_keys is not None and real not in model.request_keys:
+ problems.append(
+ f"{real}: recommended {want!r} but the strict spec does not declare it"
+ )
+ elif real not in sent or not _same(sent[real], want):
+ problems.append(
+ f"{real}: request sends {sent.get(real, '<nothing>')!r}, want {want!r}"
+ )
+ for name in case["exclude"]:
+ if name in by_name:
+ problems.append(f"{name}: offered but must not be")
+ if workflow in ("clone", "edit", "convert") and "voice" in sent:
+ problems.append("voice: sent on a workflow that speaks in a given voice")
+ assert not problems, where + "\n " + "\n ".join(problems)
+
+
+def test_the_golden_covers_every_tts_and_music_model(catalog_hub):
+ wanted = {model_id for model_id, entry in SPECS.items() if entry["task"] != "sep"}
+ assert set(GOLDEN) == wanted
+ for model_id, cases in GOLDEN.items():
+ for model in _resolved(model_id):
+ covered = {
+ c["workflow"]
+ for c in cases
+ if not c.get("variant")
+ or model.variant.key.lower().startswith(c["variant"].lower() + "/")
+ }
+ assert covered == set(model.workflows), (model_id, model.variant.key)
diff --git a/studio/backend/tests/test_audio_cpp_option_table.py b/studio/backend/tests/test_audio_cpp_option_table.py
new file mode 100644
index 00000000000..cc29270ca64
--- /dev/null
+++ b/studio/backend/tests/test_audio_cpp_option_table.py
@@ -0,0 +1,298 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Every audio.cpp model in the Audio catalog, one at a time: the options each workflow offers, the
+recommended value each one starts from, and that a request with nothing set sends those values.
+
+The specs are the ones embedded in the published GGUFs (``fixtures/audio_cpp_catalog_specs.json``),
+read with no runtime ``model_specs/``, as a prebuilt runtime runs them."""
+
+import json
+import re
+import sys
+from dataclasses import replace
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+from core.inference import audio_cpp_backend, audio_cpp_files
+from core.inference import audio_cpp_models as acm
+from core.inference.audio_cpp_option_table import (
+ AUDIO_CPP_OPTION_TABLE_TAGS,
+ OPTION_RULES,
+ STUDIO_OPTION_LISTS,
+)
+from core.inference.audio_cpp_spec_families import AUDIO_CPP_SPEC_FAMILIES_TAGS
+
+_TESTS_DIR = str(Path(__file__).resolve().parent)
+if _TESTS_DIR not in sys.path:
+ sys.path.insert(0, _TESTS_DIR)
+
+from test_audio_cpp_models import _gguf_bytes, _put, _snapshot # noqa: E402
+
+FIXTURE = Path(__file__).parent / "fixtures" / "audio_cpp_catalog_specs.json"
+CATALOG = Path(__file__).resolve().parents[2] / "frontend/src/features/audio/audio-cpp-catalog.ts"
+SPECS = json.loads(FIXTURE.read_text(encoding = "utf-8"))["models"]
+WORKFLOWS = ("speak", "clone", "edit", "convert", "music")
+
+
+def catalog_hub(tmp_path, monkeypatch):
+ """A fake HF cache holding every catalog model's files, each GGUF carrying its embedded spec."""
+ root = tmp_path / "hub"
+ root.mkdir()
+ monkeypatch.setattr(acm, "_hub_cache", lambda: root)
+ monkeypatch.setattr(audio_cpp_files, "_hub_cache", lambda: root)
+ monkeypatch.setattr(acm, "runtime_spec", lambda family: None)
+ monkeypatch.setattr(acm, "runtime_knows_family", lambda family: None)
+ monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+ acm.forget()
+ with acm._resolve_lock:
+ acm._downloaded_cache.clear()
+ snaps = {}
+ for model_id, entry in SPECS.items():
+ repo, _, folder = model_id.partition("/audio.cpp-gguf/")
+ repo = acm.AUDIO_CPP_REPO if folder else model_id
+ snap = snaps.get(repo) or snaps.setdefault(repo, _snapshot(root, repo))
+ header = _gguf_bytes(family = entry["family"], spec = entry["spec"])
+ for rel in entry["files"]:
+ path = f"{folder}/{rel}" if folder else rel
+ _put(snap, path, header if rel.endswith(".gguf") else b"{}")
+ yield root
+ acm.forget()
+
+
+def _models():
+ """(model id, variant key) for every variant of every TTS and music catalog model."""
+ out = []
+ for model_id, entry in sorted(SPECS.items()):
+ if entry["task"] == "sep":
+ continue
+ out.append(model_id)
+ return out
+
+
+def _resolved(model_id):
+ model = acm.resolve(model_id, network = False)
+ assert model is not None and model.unsupported is None, (model_id, model and model.unsupported)
+ return [model, *(model.with_variant(v) for v in model.variants if v.key != model.variant.key)]
+
+
+def _wanted(model, workflow):
+ return workflow in model.workflows
+
+
+def test_the_table_is_written_for_the_pinned_runtime():
+ assert AUDIO_CPP_OPTION_TABLE_TAGS == AUDIO_CPP_SPEC_FAMILIES_TAGS
+
+
+def test_the_fixture_covers_the_catalog():
+ source = CATALOG.read_text(encoding = "utf-8")
+ seeded = {
+ (f"{acm.AUDIO_CPP_REPO}/{folder}" if folder else repo)
+ for folder, repo in re.findall(
+ r'id: (?:folder\("([^"]+)"\)|"(audio-cpp/[^"]+)"),\s*task: "(?:tts|music|sep)"', source
+ )
+ }
+ assert seeded == set(SPECS)
+
+
+def test_tables_name_known_families():
+ for family in (*STUDIO_OPTION_LISTS, *OPTION_RULES):
+ assert family in acm.FAMILIES, family
+
+
[email protected]("model_id", _models())
+def test_each_option_starts_from_a_value_it_accepts(catalog_hub, model_id):
+ for model in _resolved(model_id):
+ for workflow in WORKFLOWS:
+ if not _wanted(model, workflow):
+ continue
+ for option in acm.workflow_options(model, workflow):
+ default = option.get("default")
+ where = (model_id, model.variant.key, workflow, option["name"])
+ if option["type"] == "bool":
+ # A switch has no "unset" look: it must show what runs.
+ assert isinstance(default, bool), where
+ if default is None:
+ continue
+ assert acm.validate_options((option,), {option["name"]: default}) == {
+ option["name"]: default
+ }, where
+
+
[email protected]("model_id", _models())
+def test_a_request_with_nothing_set_sends_the_shown_values(catalog_hub, model_id):
+ for model in _resolved(model_id):
+ for workflow in WORKFLOWS:
+ if not _wanted(model, workflow):
+ continue
+ schema = acm.workflow_options(model, workflow)
+ sent = acm.filled_options(model, workflow, {})
+ # A tool panel's own options are the panel's to send.
+ own = {o["name"] for o in model.options}
+ shown = {
+ o["name"]: o["default"]
+ for o in schema
+ if o.get("default") is not None and o["name"] != "voice" and o["name"] in own
+ }
+ if model.request_keys is not None:
+ assert set(sent) <= model.request_keys, (model_id, workflow)
+ shown = {k: v for k, v in shown.items() if k in model.request_keys}
+ assert sent == shown, (model_id, model.variant.key, workflow)
+ if workflow in ("clone", "edit", "convert"):
+ assert "voice" not in {o["name"] for o in schema}
+
+
[email protected]("model_id", _models())
+def test_user_values_win_over_recommended_ones(catalog_hub, model_id):
+ model = _resolved(model_id)[0]
+ for workflow in WORKFLOWS:
+ if not _wanted(model, workflow):
+ continue
+ schema = acm.workflow_options(model, workflow)
+ numeric = next((o for o in schema if o["type"] in ("int", "float")), None)
+ if numeric is None:
+ continue
+ value = numeric["max"] if numeric.get("max") is not None else (numeric.get("min") or 0) + 1
+ sent = acm.request_options(schema, {numeric["name"]: value}, None)
+ assert (
+ sent[numeric["name"]]
+ == acm.validate_options(schema, {numeric["name"]: value})[numeric["name"]]
+ )
+
+
+ "model_id, name",
+ [
+ ("audio-cpp/audio.cpp-gguf/VoxCPM2-GGUF", "retry_badcase_max_times"),
+ ("audio-cpp/audio.cpp-gguf/VoxCPM2-GGUF", "retry_badcase_ratio_threshold"),
+ ("audio-cpp/audio.cpp-gguf/Fish-Audio-S2-Pro-GGUF", "top_p"),
+ ],
+)
+def test_a_zero_the_runtime_refuses_is_never_sent(catalog_hub, model_id, name):
+ # audio.cpp throws on these at 0 (voxcpm2/session.cpp:670-676, fish_audio/session.cpp:136).
+ model = _resolved(model_id)[0]
+ sent = acm.filled_options(model, "speak", {name: 0})
+ assert sent[name] > 0
+
+
+def test_status_lists_each_workflow_where_it_differs(catalog_hub):
+ for model_id in _models():
+ model = _resolved(model_id)[0]
+ fields = audio_cpp_backend.model_info_fields(model)
+ primary = "speak" if "speak" in model.workflows else next(iter(model.workflows))
+ assert fields["audio_options"] == [dict(o) for o in acm.workflow_options(model, primary)]
+ by_workflow = fields["audio_options_by_workflow"] or {}
+ for workflow in model.workflows:
+ listed = by_workflow.get(workflow, fields["audio_options"])
+ assert listed == [dict(o) for o in acm.workflow_options(model, workflow)], (
+ model_id,
+ workflow,
+ )
+
+
+MAGPIE = f"{acm.AUDIO_CPP_REPO}/MagpieTTS-Multilingual-357M-GGUF"
+
+
+def _stale_magpie():
+ """MagpieTTS as resolved before its download: options read from the ``orig`` file's header,
+ which names the speaker option ``speaker`` where the served Q8_0 says ``voice_id``."""
+ served = acm.resolve(MAGPIE, network = False)
+ speaker = {
+ "name": "speaker",
+ "type": "int",
+ "description": "",
+ "required": False,
+ "default": 0,
+ "min": None,
+ "max": None,
+ "values": None,
+ }
+ own = tuple(o for o in served.spec_options if o["name"] != "voice_id")
+ keys = (served.request_keys or frozenset()) - {"voice_id"} | {"speaker"}
+ stale = replace(
+ served, options = (speaker, *own), spec_options = (speaker, *own), request_keys = keys
+ )
+ return served, stale
+
+
[email protected]("downloaded", [True, False])
+def test_a_load_that_downloads_fills_the_served_files_options(catalog_hub, monkeypatch, downloaded):
+ from core.inference import audio_cpp_server
+
+ served, stale = _stale_magpie()
+ backend = audio_cpp_backend.AudioCppBackend()
+ monkeypatch.setattr(backend, "_ensure_downloaded", lambda model, token: downloaded)
+ monkeypatch.setattr(backend, "_start_server", lambda model, *a, **k: None)
+ monkeypatch.setattr(audio_cpp_server, "model_runtime_problem", lambda model: None)
+ backend.load_model(SimpleNamespace(identifier = MAGPIE, audio_cpp = stale))
+ filled = acm.filled_options(backend._model, "speak", {})
+ if downloaded:
+ assert "voice_id" in filled and "speaker" not in filled
+ assert filled == acm.filled_options(served, "speak", {})
+ else:
+ # Nothing was fetched, so the files on disk did not change: the model is kept as given.
+ assert "speaker" in filled
+
+
+def test_each_cached_quantization_offers_its_own_files_options(catalog_hub):
+ """With both MagpieTTS files cached, each variant's options come from the file it serves."""
+ q8 = SPECS[MAGPIE]["spec"]
+ orig = json.loads(json.dumps(q8))
+ for option in orig["options"]["request"]:
+ if option["name"] == "voice_id":
+ option["name"] = "speaker"
+ snap = _snapshot(catalog_hub)
+ folder = "MagpieTTS-Multilingual-357M-GGUF"
+ _put(
+ snap,
+ f"{folder}/magpie-tts-multilingual-357m-orig.gguf",
+ _gguf_bytes(family = "magpie_tts", spec = orig),
+ )
+ _put(
+ snap,
+ f"{folder}/magpie-tts-multilingual-357m-q8_0.gguf",
+ _gguf_bytes(family = "magpie_tts", spec = q8),
+ )
+ acm.forget()
+ served = acm.resolve(MAGPIE, network = False)
+ assert served.variant.key == "Q8_0"
+ filled = acm.filled_options(served, "speak", {})
+ assert "voice_id" in filled and "speaker" not in filled
+ assert "speaker" not in (served.request_keys or ())
+ acm.forget()
+ other = acm.resolve(MAGPIE, "orig", network = False)
+ assert "speaker" in acm.filled_options(other, "speak", {})
+
+
+ACE_STEP = f"{acm.AUDIO_CPP_REPO}/ACE-Step1.5-GGUF"
+
+
+ "name, base",
+ [
+ ("ace-step-1.5-base-q8_0.gguf", True),
+ ("acestep-v15-xl-sft-q8_0.gguf", True),
+ ("ace-step-1.5-turbo-q8_0.gguf", False),
+ # The runtime reads the file name only, so a base/ folder alone is turbo.
+ ("base/model-q8_0.gguf", False),
+ ("ace step base v1.gguf", True),
+ ],
+)
+def test_ace_step_variant_follows_the_file_name_like_the_runtime(catalog_hub, tmp_path, name, base):
+ """audio.cpp picks base/SFT from "base" or "sft" in the GGUF name, wherever it is stored."""
+ path = tmp_path / name
+ path.parent.mkdir(parents = True, exist_ok = True)
+ path.write_bytes(_gguf_bytes(family = "ace_step", spec = SPECS[ACE_STEP]["spec"]))
+ model = acm.resolve(str(path), network = False)
+ options = {o["name"]: o for o in acm.workflow_options(model, "music")}
+ filled = acm.filled_options(model, "music", {})
+ if base:
+ assert options["num_inference_steps"]["max"] == 200
+ assert filled["num_inference_steps"] == 32 and filled["guidance_scale"] == 7.0
+ assert filled["shift"] == 1.0
+ else:
+ assert options["num_inference_steps"]["max"] == 8 and "guidance_scale" not in options
+ assert filled["shift"] == 3.0
diff --git a/studio/frontend/src/features/audio/audio-options-fields.tsx b/studio/frontend/src/features/audio/audio-options-fields.tsx
index bf952dd7fe3..773cf51e339 100644
--- a/studio/frontend/src/features/audio/audio-options-fields.tsx
+++ b/studio/frontend/src/features/audio/audio-options-fields.tsx
@@ -18,6 +18,7 @@ import {
audioOptionDisplayValue,
audioOptionFloatStep,
audioOptionLabel,
+ audioOptionUsesSlider,
audioVoiceLabel,
coerceAudioOptionValue,
} from "./audio-options";
@@ -29,7 +30,7 @@ function OptionHint({ text }: { text?: string | null }) {
}
/** One control per declared option, drawn from the type alone: a switch, a slider when the
- * range is bounded (else a number box), a select, or a text box. */
+ * range is bounded and has a default (else a number box), a select, or a text box. */
export function AudioOptionFields({
specs,
values,
@@ -70,12 +71,7 @@ export function AudioOptionFields({
</div>
);
}
- if (
- (spec.type === "int" || spec.type === "float") &&
- spec.min != null &&
- spec.max != null &&
- spec.max > spec.min
- ) {
+ if (audioOptionUsesSlider(spec) && spec.min != null && spec.max != null) {
const step = spec.type === "int" ? 1 : audioOptionFloatStep(spec.min, spec.max);
return (
<ParamSlider
diff --git a/studio/frontend/src/features/audio/audio-options.ts b/studio/frontend/src/features/audio/audio-options.ts
index b437637441f..74c6232e571 100644
--- a/studio/frontend/src/features/audio/audio-options.ts
+++ b/studio/frontend/src/features/audio/audio-options.ts
@@ -65,6 +65,35 @@ export function parseAudioOptions(raw: unknown): AudioOptionSpec[] {
return specs;
}
+/** The options one workflow of the loaded model takes: its own list where the status gives one
+ * (`audio_options_by_workflow`), else the primary list (`audio_options`). */
+export function audioOptionSpecsFor(
+ status:
+ | {
+ audio_options?: unknown;
+ audio_options_by_workflow?: Record<string, unknown> | null;
+ }
+ | null
+ | undefined,
+ workflow: string,
+): AudioOptionSpec[] {
+ return parseAudioOptions(
+ status?.audio_options_by_workflow?.[workflow] ?? status?.audio_options,
+ );
+}
+
+/** A slider needs a range and a value to sit on: without a default it would rest at the
+ * minimum and look like a setting, so such an option gets a number box instead. */
+export function audioOptionUsesSlider(spec: AudioOptionSpec): boolean {
+ return (
+ (spec.type === "int" || spec.type === "float") &&
+ spec.min != null &&
+ spec.max != null &&
+ spec.max > spec.min &&
+ spec.default != null
+ );
+}
+
/** `value` as this option accepts it, or undefined when it cannot be: numbers are clamped to the
* declared range and ints rounded, an enum takes only its listed values. */
export function coerceAudioOptionValue(
@@ -101,8 +130,8 @@ export function audioOptionDisplayValue(
return spec.default ?? undefined;
}
-/** The values to send: only ones the user set and the schema still accepts. Everything else is
- * left to the model's own defaults on the server. */
+/** The values to send: only ones the user set and the schema still accepts. The server fills
+ * every other option with the same recommended default the controls show. */
export function audioOptionsForRequest(
specs: readonly AudioOptionSpec[],
values: AudioOptionValues,
diff --git a/studio/frontend/src/features/audio/audio-page.tsx b/studio/frontend/src/features/audio/audio-page.tsx
index b47eefcdeec..1fbbe8086a8 100644
--- a/studio/frontend/src/features/audio/audio-page.tsx
+++ b/studio/frontend/src/features/audio/audio-page.tsx
@@ -655,7 +655,6 @@ export function AudioPage({
setSelectedId,
pendingTranscribeRelease,
replayQueuedTtsPick,
- audioOptionSpecs,
audioOptionValues,
sttRepo: cloneSttRepo,
});
@@ -676,7 +675,6 @@ export function AudioPage({
setSelectedId,
pendingTranscribeRelease,
replayQueuedTtsPick,
- audioOptionSpecs,
audioOptionValues,
sttRepo: cloneSttRepo,
});
@@ -697,7 +695,6 @@ export function AudioPage({
setSelectedId,
pendingTranscribeRelease,
replayQueuedTtsPick,
- audioOptionSpecs,
audioOptionValues,
sttRepo: cloneSttRepo,
});
@@ -1921,13 +1918,19 @@ export function AudioPage({
<AudioActiveProvider value={active}>
<CloneRail
{...railProps}
+ audioOptionSpecs={clone.audioOptionSpecs}
clone={clone}
historyClips={clips}
/>
</AudioActiveProvider>
) : ttsWorkflow === "edit" ? (
<AudioActiveProvider value={active}>
- <EditRail {...railProps} edit={edit} historyClips={clips} />
+ <EditRail
+ {...railProps}
+ audioOptionSpecs={edit.audioOptionSpecs}
+ edit={edit}
+ historyClips={clips}
+ />
</AudioActiveProvider>
) : ttsWorkflow === "convert" ? (
<AudioActiveProvider value={active}>
diff --git a/studio/frontend/src/features/audio/hooks/use-clone-generation.ts b/studio/frontend/src/features/audio/hooks/use-clone-generation.ts
index 2a0736c22ab..3ab6e74c4ca 100644
--- a/studio/frontend/src/features/audio/hooks/use-clone-generation.ts
+++ b/studio/frontend/src/features/audio/hooks/use-clone-generation.ts
@@ -10,9 +10,9 @@ import {
runAudio,
} from "../api";
import {
- type AudioOptionSpec,
type AudioOptionValues,
audioOptionLabel,
+ audioOptionSpecsFor,
audioOptionsForRequest,
missingRequiredAudioOptions,
} from "../audio-options";
@@ -119,7 +119,6 @@ export function useCloneGeneration({
setSelectedId,
pendingTranscribeRelease,
replayQueuedTtsPick,
- audioOptionSpecs,
audioOptionValues,
sttRepo,
}: Pick<
@@ -140,7 +139,6 @@ export function useCloneGeneration({
"refreshGallery" | "selectClip" | "setFallbackClip" | "setSelectedId"
> &
Pick<AudioModelSlot, "pendingTranscribeRelease" | "replayQueuedTtsPick"> & {
- audioOptionSpecs: AudioOptionSpec[];
audioOptionValues: AudioOptionValues;
sttRepo: string | null;
}) {
@@ -150,6 +148,10 @@ export function useCloneGeneration({
const language = useAudioCloneStore((state) => state.language);
const storedToolValues = useAudioCloneStore((state) => state.toolValues);
const model = status?.active_model ?? null;
+ const audioOptionSpecs = useMemo(
+ () => audioOptionSpecsFor(status, "clone"),
+ [status],
+ );
const [generationError, setGenerationError] = useState<string | null>(null);
const [referenceStatus, setReferenceStatus] = useState<AudioSourceStatus>({
phase: "idle",
@@ -445,6 +447,7 @@ export function useCloneGeneration({
toolValues,
handleToolValueChange,
claimedOptions,
+ audioOptionSpecs,
advancedOptionSpecs,
blocker,
inputsReady: inputBlocker === null,
diff --git a/studio/frontend/src/features/audio/hooks/use-convert-generation.ts b/studio/frontend/src/features/audio/hooks/use-convert-generation.ts
index 9b015eae725..a75a49478f9 100644
--- a/studio/frontend/src/features/audio/hooks/use-convert-generation.ts
+++ b/studio/frontend/src/features/audio/hooks/use-convert-generation.ts
@@ -5,12 +5,11 @@ import { toast } from "@/lib/toast";
import { useCallback, useEffect, useMemo, useRef, useState } from "react";
import { AudioApiError, runAudio } from "../api";
import {
- type AudioOptionSpec,
type AudioOptionValues,
audioOptionLabel,
+ audioOptionSpecsFor,
audioOptionsForRequest,
missingRequiredAudioOptions,
- parseAudioOptions,
} from "../audio-options";
import {
type AudioSourceSelection,
@@ -76,7 +75,6 @@ export function useConvertGeneration({
setSelectedId,
pendingTranscribeRelease,
replayQueuedTtsPick,
- audioOptionSpecs,
audioOptionValues,
sttRepo,
}: Pick<
@@ -97,7 +95,6 @@ export function useConvertGeneration({
"refreshGallery" | "selectClip" | "setFallbackClip" | "setSelectedId"
> &
Pick<AudioModelSlot, "pendingTranscribeRelease" | "replayQueuedTtsPick"> & {
- audioOptionSpecs: AudioOptionSpec[];
audioOptionValues: AudioOptionValues;
sttRepo: string | null;
}) {
@@ -110,6 +107,10 @@ export function useConvertGeneration({
const sourceText = useAudioConvertStore((state) => state.sourceText);
const storedToolValues = useAudioCloneStore((state) => state.toolValues);
const model = status?.active_model ?? null;
+ const audioOptionSpecs = useMemo(
+ () => audioOptionSpecsFor(status, "convert"),
+ [status],
+ );
const [generationError, setGenerationError] = useState<string | null>(null);
const [sourceStatus, setSourceStatus] = useState<AudioSourceStatus>(IDLE);
const [targetStatus, setTargetStatus] = useState<AudioSourceStatus>(IDLE);
@@ -185,12 +186,6 @@ export function useConvertGeneration({
.setToolValue(toolValueKey(model, "convert", panelId), value),
[model],
);
- const convertOptionSpecs = useMemo(() => {
- const byWorkflow = status?.audio_options_by_workflow?.convert;
- return byWorkflow === undefined
- ? audioOptionSpecs
- : parseAudioOptions(byWorkflow);
- }, [status?.audio_options_by_workflow, audioOptionSpecs]);
const toolRequest = useMemo(
() =>
collectToolRequest(
@@ -198,7 +193,7 @@ export function useConvertGeneration({
toolValues,
{ text: "", referenceText: sourceText, hasReference: source !== null },
{ ...toolContext, convertMode: mode },
- convertOptionSpecs,
+ audioOptionSpecs,
),
[
toolPanels,
@@ -207,7 +202,7 @@ export function useConvertGeneration({
source,
toolContext,
mode,
- convertOptionSpecs,
+ audioOptionSpecs,
],
);
const style = effectiveConvertStyle(
@@ -222,8 +217,8 @@ export function useConvertGeneration({
[toolPanels],
);
const advancedOptionSpecs = useMemo(
- () => convertOptionSpecs.filter((spec) => !claimedOptions.has(spec.name)),
- [convertOptionSpecs, claimedOptions],
+ () => audioOptionSpecs.filter((spec) => !claimedOptions.has(spec.name)),
+ [audioOptionSpecs, claimedOptions],
);
const nextTask = convertServerTask(caps, status?.audio_workflow_tasks, mode);
diff --git a/studio/frontend/src/features/audio/hooks/use-edit-generation.ts b/studio/frontend/src/features/audio/hooks/use-edit-generation.ts
index cbc44bd67c5..4d0c8bd2885 100644
--- a/studio/frontend/src/features/audio/hooks/use-edit-generation.ts
+++ b/studio/frontend/src/features/audio/hooks/use-edit-generation.ts
@@ -5,9 +5,9 @@ import { toast } from "@/lib/toast";
import { useCallback, useEffect, useMemo, useRef, useState } from "react";
import { AudioApiError, runAudio } from "../api";
import {
- type AudioOptionSpec,
type AudioOptionValues,
audioOptionLabel,
+ audioOptionSpecsFor,
audioOptionsForRequest,
missingRequiredAudioOptions,
} from "../audio-options";
@@ -73,7 +73,6 @@ export function useEditGeneration({
setSelectedId,
pendingTranscribeRelease,
replayQueuedTtsPick,
- audioOptionSpecs,
audioOptionValues,
sttRepo,
}: Pick<
@@ -94,7 +93,6 @@ export function useEditGeneration({
"refreshGallery" | "selectClip" | "setFallbackClip" | "setSelectedId"
> &
Pick<AudioModelSlot, "pendingTranscribeRelease" | "replayQueuedTtsPick"> & {
- audioOptionSpecs: AudioOptionSpec[];
audioOptionValues: AudioOptionValues;
sttRepo: string | null;
}) {
@@ -106,6 +104,10 @@ export function useEditGeneration({
const delivery = useAudioEditStore((state) => state.delivery);
const storedToolValues = useAudioCloneStore((state) => state.toolValues);
const model = status?.active_model ?? null;
+ const audioOptionSpecs = useMemo(
+ () => audioOptionSpecsFor(status, "edit"),
+ [status],
+ );
const [generationError, setGenerationError] = useState<string | null>(null);
const [phaseLabel, setPhaseLabel] = useState<string | null>(null);
const [sourceStatus, setSourceStatus] = useState<AudioSourceStatus>({
@@ -407,6 +409,7 @@ export function useEditGeneration({
toolValues,
handleToolValueChange,
claimedOptions,
+ audioOptionSpecs,
advancedOptionSpecs,
core,
blocker,
diff --git a/studio/frontend/src/features/audio/hooks/use-speech-generation.ts b/studio/frontend/src/features/audio/hooks/use-speech-generation.ts
index 2cbff61fa7f..527a0ed9610 100644
--- a/studio/frontend/src/features/audio/hooks/use-speech-generation.ts
+++ b/studio/frontend/src/features/audio/hooks/use-speech-generation.ts
@@ -8,11 +8,11 @@ import { generateAudio, runAudio } from "../api";
import { TTS_MAX_TOKENS } from "../audio-workspace-constants";
import {
audioOptionLabel,
+ audioOptionSpecsFor,
audioOptionsForRequest,
type AudioOptionValue,
type AudioOptionValues,
missingRequiredAudioOptions,
- parseAudioOptions,
readAudioOptionValues,
saveAudioOptionValues,
} from "../audio-options";
@@ -188,8 +188,8 @@ export function useSpeechGeneration({
);
const samplingControls = audioSamplingControlsApply(status?.audio_type);
const audioOptionSpecs = useMemo(
- () => parseAudioOptions(status?.audio_options),
- [status?.audio_options],
+ () => audioOptionSpecsFor(status, workflow),
+ [status, workflow],
);
const audioOptionsModel = status?.active_model ?? null;
const [audioOptionValues, setAudioOptionValues] = useState<AudioOptionValues>(
diff --git a/studio/frontend/src/features/audio/tools/panel-logic.ts b/studio/frontend/src/features/audio/tools/panel-logic.ts
index 3183a54cb34..27d612e33aa 100644
--- a/studio/frontend/src/features/audio/tools/panel-logic.ts
+++ b/studio/frontend/src/features/audio/tools/panel-logic.ts
@@ -3,6 +3,7 @@
// JSX-free so the node test runner can load it; clone-panels.tsx / speak-panels.tsx add Components.
+import type { AudioOptionSpec } from "../audio-options";
import { type AudioSourceSelection, sourceRefOf } from "../audio-run-request";
import { INDEX_TTS2_EMOTIONS, emotionVectorString } from "../clone-policy";
import type { AudioSourceStatus } from "../hooks/audio-source-state";
@@ -13,6 +14,25 @@ type AudioToolPanelLogic<V> = Omit<AudioToolPanel<V>, "Component">;
const clamp = (x: number, lo: number, hi: number) =>
Math.min(hi, Math.max(lo, x));
+// A panel starts where the server would: the option's declared default, which it fills when unset.
+function numberDefault(
+ specs: readonly AudioOptionSpec[],
+ name: string,
+ fallback: number,
+): number {
+ const value = specs.find((spec) => spec.name === name)?.default;
+ return typeof value === "number" && Number.isFinite(value) ? value : fallback;
+}
+
+function boolDefault(
+ specs: readonly AudioOptionSpec[],
+ name: string,
+ fallback: boolean,
+): boolean {
+ const value = specs.find((spec) => spec.name === name)?.default;
+ return typeof value === "boolean" ? value : fallback;
+}
+
export interface TimbreOnlyValue {
timbreOnly: boolean;
}
@@ -23,7 +43,9 @@ export const qwen3TimbreLogic: AudioToolPanelLogic<TimbreOnlyValue> = {
workflows: ["clone"],
title: "Timbre only",
claims: ["x_vector_only_mode"],
- initial: () => ({ timbreOnly: false }),
+ initial: (specs) => ({
+ timbreOnly: boolDefault(specs, "x_vector_only_mode", false),
+ }),
toRequest: (value) =>
value.timbreOnly
? {
@@ -68,11 +90,11 @@ export const indexTts2EmotionLogic: AudioToolPanelLogic<EmotionValue> = {
"use_emotion_text",
"emotion_text",
],
- initial: () => ({
+ initial: (specs) => ({
mode: "text",
text: "",
vector: INDEX_TTS2_EMOTIONS.map(() => 0),
- alpha: 0.7,
+ alpha: numberDefault(specs, "emotion_alpha", 0.7),
source: null,
}),
toRequest: (value): AudioRunPatch => {
@@ -124,7 +146,10 @@ export const chatterboxExpressivenessLogic: AudioToolPanelLogic<ExpressivenessVa
workflows: ["clone"],
title: "Expressiveness",
claims: ["exaggeration", "guidance_scale"],
- initial: () => ({ exaggeration: 0.5, guidance: 0.5 }),
+ initial: (specs) => ({
+ exaggeration: numberDefault(specs, "exaggeration", 0.5),
+ guidance: numberDefault(specs, "guidance_scale", 0.5),
+ }),
toRequest: (value) => ({
options: {
exaggeration: clamp(value.exaggeration, 0, 2),
diff --git a/studio/frontend/tests/audio-clone-panels.test.ts b/studio/frontend/tests/audio-clone-panels.test.ts
index e4a7ea3db96..d14c2a530b7 100644
--- a/studio/frontend/tests/audio-clone-panels.test.ts
+++ b/studio/frontend/tests/audio-clone-panels.test.ts
@@ -156,6 +156,33 @@ test("a kept value is laid over the panel's defaults", () => {
);
});
+test("clone panels start at the option defaults the server fills", () => {
+ const specs = [
+ { name: "emotion_alpha", type: "float" as const, min: 0, max: 1, default: 1 },
+ { name: "exaggeration", type: "float" as const, min: 0, max: 2, default: 0.8 },
+ { name: "guidance_scale", type: "float" as const, min: 0, max: 5, default: 0.3 },
+ { name: "x_vector_only_mode", type: "bool" as const, default: true },
+ ];
+ assert.equal(indexTts2EmotionLogic.initial(specs).alpha, 1);
+ assert.deepEqual(chatterboxExpressivenessLogic.initial(specs), {
+ exaggeration: 0.8,
+ guidance: 0.3,
+ });
+ assert.deepEqual(qwen3TimbreLogic.initial(specs), { timbreOnly: true });
+ // Without a declared default each keeps its old starting point.
+ assert.equal(indexTts2EmotionLogic.initial([]).alpha, 0.7);
+ assert.deepEqual(chatterboxExpressivenessLogic.initial([]), {
+ exaggeration: 0.5,
+ guidance: 0.5,
+ });
+ assert.deepEqual(qwen3TimbreLogic.initial([]), { timbreOnly: false });
+ // A kept strength still wins over the default.
+ assert.equal(
+ panelValue(indexTts2EmotionLogic, { "index-emotion": { alpha: 0.4 } }, specs).alpha,
+ 0.4,
+ );
+});
+
test("the IndexTTS2 mixer sends eight weights and the strength", () => {
const vector = [0.8, 0, 0, 0, 0, 0, 0.2, 0];
assert.equal(emotionVectorString(vector), "0.8,0,0,0,0,0,0.2,0");
diff --git a/studio/frontend/tests/audio-cpp-catalog.test.ts b/studio/frontend/tests/audio-cpp-catalog.test.ts
index 2f3fb3c6190..cb084bcfc2a 100644
--- a/studio/frontend/tests/audio-cpp-catalog.test.ts
+++ b/studio/frontend/tests/audio-cpp-catalog.test.ts
@@ -793,7 +793,7 @@ test("GGUF runtime speech trades sampling controls for the model's own options",
/\{musicGeneration \|\|\s*samplingControls \|\|\s*audioOptionSpecs\.length > 0 \? \(\s*<AdvancedDisclosure/,
);
assert.match(page, /!musicGeneration && samplingControls && temperatureEdited/);
- assert.match(page, /parseAudioOptions\(status\?\.audio_options\)/);
+ assert.match(page, /audioOptionSpecsFor\(status, workflow\)/);
assert.match(page, /\{ audio_options: requestOptions \}/);
assert.match(page, /"Voice or style description"/);
});
diff --git a/studio/frontend/tests/audio-music-panels.test.ts b/studio/frontend/tests/audio-music-panels.test.ts
index 47b383c9252..c2adc09da36 100644
--- a/studio/frontend/tests/audio-music-panels.test.ts
+++ b/studio/frontend/tests/audio-music-panels.test.ts
@@ -157,6 +157,19 @@ test("Stable Audio's sampler sends pingpong or euler and whole steps", () => {
);
});
+test("the steps range follows the declared range and recommended default", () => {
+ assert.deepEqual(
+ stepsRange([
+ { name: "num_inference_steps", type: "int", min: 1, max: 200, default: 32 },
+ ]),
+ { min: 1, max: 200, default: 32 },
+ );
+ assert.deepEqual(
+ stepsRange([{ name: "num_inference_steps", type: "int", min: 1, max: 8, default: 8 }]),
+ { min: 1, max: 8, default: 8 },
+ );
+});
+
test("the Music studio's Advanced hides speech sampling it never sends", () => {
const page = readSrc("features/audio/pages/music-page.tsx");
assert.match(
diff --git a/studio/frontend/tests/audio-options.test.ts b/studio/frontend/tests/audio-options.test.ts
index 2e2964c7db7..11dd5e82143 100644
--- a/studio/frontend/tests/audio-options.test.ts
+++ b/studio/frontend/tests/audio-options.test.ts
@@ -8,7 +8,7 @@
import assert from "node:assert/strict";
import test from "node:test";
-import { installLocalStorageFake } from "./helpers/kit.ts";
+import { installLocalStorageFake, readSrc } from "./helpers/kit.ts";
installLocalStorageFake();
const {
@@ -16,6 +16,8 @@ const {
audioOptionDisplayValue,
audioOptionFloatStep,
audioOptionLabel,
+ audioOptionSpecsFor,
+ audioOptionUsesSlider,
audioOptionsForRequest,
audioVoiceLabel,
coerceAudioOptionValue,
@@ -159,3 +161,111 @@ test("built-in voices read as names, and Kokoro ids also name their language", (
assert.equal(audioVoiceLabel("af_heart", "kitten_tts"), "Af Heart");
assert.equal(audioVoiceLabel("Custom Voice.wav", null), "Custom Voice.wav");
});
+
+test("each workflow reads its own option list, else the primary one", () => {
+ const status = {
+ audio_options: [{ name: "voice", type: "enum", values: ["a", "b"], default: "a" }],
+ audio_options_by_workflow: {
+ clone: [{ name: "emotion_alpha", type: "float", min: 0, max: 1, default: 1 }],
+ convert: [],
+ },
+ };
+ assert.deepEqual(
+ audioOptionSpecsFor(status, "speak").map((spec) => spec.name),
+ ["voice"],
+ );
+ assert.deepEqual(
+ audioOptionSpecsFor(status, "clone").map((spec) => spec.name),
+ ["emotion_alpha"],
+ );
+ // An empty list is the workflow's own answer, not a gap to fill from the primary list.
+ assert.deepEqual(audioOptionSpecsFor(status, "convert"), []);
+ assert.deepEqual(
+ audioOptionSpecsFor(status, "edit").map((spec) => spec.name),
+ ["voice"],
+ );
+ assert.deepEqual(
+ audioOptionSpecsFor({ ...status, audio_options_by_workflow: null }, "clone").map(
+ (spec) => spec.name,
+ ),
+ ["voice"],
+ );
+ assert.deepEqual(audioOptionSpecsFor(null, "speak"), []);
+
+ // Every generation hook picks its own workflow's list.
+ const hooks = {
+ speech: "workflow",
+ clone: '"clone"',
+ edit: '"edit"',
+ convert: '"convert"',
+ } as const;
+ for (const [hook, workflow] of Object.entries(hooks)) {
+ const source = readSrc(`features/audio/hooks/use-${hook}-generation.ts`);
+ assert.ok(
+ source.includes(`audioOptionSpecsFor(status, ${workflow})`),
+ `use-${hook}-generation reads its own list`,
+ );
+ assert.ok(!source.includes("parseAudioOptions("), `use-${hook}-generation`);
+ }
+ // Clone and Edit rails show the list their requests are validated against.
+ const page = readSrc("features/audio/audio-page.tsx");
+ assert.match(page, /audioOptionSpecs=\{clone\.audioOptionSpecs\}/);
+ assert.match(page, /audioOptionSpecs=\{edit\.audioOptionSpecs\}/);
+});
+
+test("a slider only shows for a bounded number with a default", () => {
+ const [steps, guidance, free, flag] = parseAudioOptions([
+ { name: "steps", type: "int", min: 1, max: 100, default: 8 },
+ { name: "guidance", type: "float", min: 0, max: 10 },
+ { name: "shift", type: "float", min: 1, default: 3 },
+ { name: "flag", type: "bool", default: false },
+ ]);
+ assert.equal(audioOptionUsesSlider(steps), true);
+ // No default: a number box with "Model default", not a slider resting at its minimum.
+ assert.equal(audioOptionUsesSlider(guidance), false);
+ assert.equal(audioOptionUsesSlider(free), false);
+ assert.equal(audioOptionUsesSlider(flag), false);
+ assert.equal(audioOptionDisplayValue(guidance, {}), undefined);
+ const fields = readSrc("features/audio/audio-options-fields.tsx");
+ assert.match(fields, /audioOptionUsesSlider\(spec\)/);
+ assert.match(fields, /"Model default"/);
+});
+
+test("Reset model options returns every control to the recommended value", () => {
+ const specs = parseAudioOptions(YUE2_OPTIONS);
+ const model = "audio-cpp/Yue2-3B-GGUF";
+ saveAudioOptionValues(model, {
+ num_inference_steps: 120,
+ guidance_scale: 4,
+ cot: "full",
+ abc: true,
+ });
+ assert.deepEqual(Object.keys(audioOptionsForRequest(specs, readAudioOptionValues(model))), [
+ "num_inference_steps",
+ "guidance_scale",
+ "cot",
+ "abc",
+ ]);
+ // What the reset handler does: forget the model's stored values.
+ saveAudioOptionValues(model, {});
+ const values = readAudioOptionValues(model);
+ assert.deepEqual(values, {});
+ assert.deepEqual(
+ specs.map((spec) => audioOptionDisplayValue(spec, values)),
+ specs.map((spec) => spec.default ?? undefined),
+ );
+ assert.deepEqual(
+ specs.map((spec) => audioOptionDisplayValue(spec, values)),
+ [50, 1.5, "off", false, undefined, undefined],
+ );
+ // Nothing user-set is sent, so the server fills the same recommended defaults.
+ assert.deepEqual(audioOptionsForRequest(specs, values), {});
+ const raw = globalThis.localStorage.getItem(AUDIO_OPTIONS_STORAGE_KEY);
+ assert.ok(!raw || !(JSON.parse(raw) as Record<string, unknown>)["audio-cpp/yue2-3b-gguf"]);
+
+ const hook = readSrc("features/audio/hooks/use-speech-generation.ts");
+ assert.match(
+ hook,
+ /handleAudioOptionsReset = useCallback\(\(\) => \{\s*setAudioOptionValues\(\{\}\);\s*saveAudioOptionValues\(audioOptionsModel, \{\}\);/,
+ );
+});