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: send the TTS language to Supertonic and add a Speak language picker (#14511)

Daniel Han32491ae49859
diff --git a/studio/backend/core/inference/audio_cpp_backend.py b/studio/backend/core/inference/audio_cpp_backend.py
index 503e67fb71e..148f25e454f 100644
--- a/studio/backend/core/inference/audio_cpp_backend.py
+++ b/studio/backend/core/inference/audio_cpp_backend.py
@@ -204,9 +204,37 @@ def _reread_after_download(model: AudioCppModel, hf_token: Optional[str]) -> Aud
return fresh
-# Strict specs that list no language option yet read the request's top-level language
-# (audio.cpp extract_request_language: VoxCPM1 keeps Cantonese text Traditional with zh-HK).
-_READS_TOP_LEVEL_LANGUAGE = frozenset({"voxcpm1", "audio8_tts"})
+# Families that read the request's top-level language, not an option (audio.cpp extract_request_language:
+# VoxCPM1 keeps Cantonese text Traditional with zh-HK; Supertonic generation_options_from_request).
+_READS_TOP_LEVEL_LANGUAGE = frozenset({"voxcpm1", "audio8_tts", "supertonic"})
+# Supertonic's tokenizer refuses anything but a bare ISO code (src/models/supertonic/tokenizer_text.cpp).
+_LANGUAGE_CODE_FAMILIES = frozenset({"supertonic"})
+# English names of the languages Supertonic reads (model_specs/supertonic.json).
+_CODE_LANGUAGE_NAMES = {
+ **_LANGUAGE_NAMES,
+ "ar": "Arabic",
+ "bg": "Bulgarian",
+ "cs": "Czech",
+ "da": "Danish",
+ "el": "Greek",
+ "et": "Estonian",
+ "fi": "Finnish",
+ "hi": "Hindi",
+ "hr": "Croatian",
+ "hu": "Hungarian",
+ "id": "Indonesian",
+ "lt": "Lithuanian",
+ "lv": "Latvian",
+ "nl": "Dutch",
+ "pl": "Polish",
+ "ro": "Romanian",
+ "sk": "Slovak",
+ "sl": "Slovenian",
+ "sv": "Swedish",
+ "tr": "Turkish",
+ "uk": "Ukrainian",
+ "vi": "Vietnamese",
+}
def model_info_fields(model: AudioCppModel) -> dict[str, Any]:
@@ -344,6 +372,17 @@ def language_name(language: Optional[str]) -> Optional[str]:
return None
+def language_code(language: Optional[str]) -> Optional[str]:
+ """``de-DE`` or ``German`` -> ``de``; None for Auto."""
+ text = str(language or "").strip()
+ if not text or text.lower() == "auto":
+ return None
+ for code, name in _CODE_LANGUAGE_NAMES.items():
+ if name.lower() == text.lower():
+ return code
+ return text.lower().replace("_", "-").split("-", 1)[0]
+
+
def speech_input(model: AudioCppModel, text: str) -> str:
"""VibeVoice refuses a script with no ``Speaker N:`` line."""
if model.family == "vibevoice" and not _SPEAKER_LINE_RE.search(text or ""):
@@ -973,11 +1012,20 @@ class AudioCppBackend:
"audio.cpp: %s declares no instruction option; not sending the description",
model.family,
)
+ top_language = None
if language and str(language).strip():
- if keys is None or "language" in keys:
+ if model.family in _READS_TOP_LEVEL_LANGUAGE:
+ top_language = (
+ language_code(language)
+ if model.family in _LANGUAGE_CODE_FAMILIES
+ else str(language).strip()
+ )
+ elif keys is None or "language" in keys:
requested["language"] = str(language).strip()
else:
logger.info("audio.cpp: %s does not declare language; not sending it", model.family)
+ if top_language:
+ body["language"] = top_language
merged = {**default_options, **requested}
if merged:
body["options"] = merged
@@ -993,16 +1041,17 @@ class AudioCppBackend:
except AudioCppRequestError as exc:
# 503 is a busy model, not a refused option. The server answers an option a family lacks with a
# 4xx or a 500 depending on where the family validates it, so both count as a refusal here.
- if not requested or exc.status == 503 or not 400 <= exc.status < 600:
+ if not (requested or top_language) or exc.status == 503 or not 400 <= exc.status < 600:
raise
# The request's own description and language are hints; the package's defaults and the
# user's options are not.
logger.info(
"audio.cpp: %s rejected %s (%s); retrying without them",
model.family,
- sorted(requested),
+ sorted([*requested, *(["language"] if top_language else [])]),
exc.detail,
)
+ body.pop("language", None)
if default_options:
body["options"] = default_options
else:
diff --git a/studio/backend/tests/test_audio_cpp_models.py b/studio/backend/tests/test_audio_cpp_models.py
index b288778cf31..db2e9aca103 100644
--- a/studio/backend/tests/test_audio_cpp_models.py
+++ b/studio/backend/tests/test_audio_cpp_models.py
@@ -1542,6 +1542,45 @@ def test_audio_options_reach_the_request_validated():
}
+def test_supertonic_gets_the_language_top_level_as_a_code():
+ # Supertonic reads only the top-level language (audio.cpp supertonic/session.cpp), never options.
+ for keys in (None, frozenset({"speed", "num_inference_steps", "text_chunk_size", "seed"})):
+ model = _speech_model("supertonic", request_keys = keys)
+ b, srv = _backend_with(model, SPEECH_WAV)
+ for given, sent in (
+ ("de", "de"),
+ (" de-DE ", "de"),
+ ("German", "de"),
+ ("hindi", "hi"),
+ ("auto", None),
+ ):
+ b.generate_audio_response("Hallo", language = given)
+ body = srv.calls[-1][1]
+ assert body.get("language") == sent and "language" not in body.get("options", {})
+ if keys is not None:
+ assert audio_cpp_backend.model_info_fields(model)["audio_request_inputs"] == [
+ "language"
+ ]
+
+
+def test_a_refused_top_level_language_is_retried_without_it():
+ from core.inference.audio_cpp_server import AudioCppRequestError
+
+ b, srv = _backend_with(_speech_model("supertonic"), SPEECH_WAV)
+ calls = []
+
+ def post_json(path, payload, **kwargs):
+ calls.append(json.loads(json.dumps(payload)))
+ if len(calls) == 1:
+ raise AudioCppRequestError(500, "invalid Supertonic language: xx")
+ return SPEECH_WAV
+
+ srv.post_json = post_json
+ b.generate_audio_response("hi", language = "xx")
+ assert calls[0]["language"] == "xx" and "language" not in calls[1]
+ assert calls[1]["voice"] == "M1"
+
+
def test_speech_retry_keeps_package_defaults_and_user_options_and_covers_500():
from core.inference.audio_cpp_server import AudioCppRequestError
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 809e6217326..fdf0fa6c9cd 100644
--- a/studio/frontend/src/features/audio/hooks/use-speech-generation.ts
+++ b/studio/frontend/src/features/audio/hooks/use-speech-generation.ts
@@ -375,7 +375,10 @@ export function useSpeechGeneration({
};
const savedVoice =
workflow === "speak" ? toolRequest.patch.inputs?.reference : undefined;
- const language = audioLanguage.trim();
+ // MOSS Local's language rides its Style panel; Supertonic's has its own.
+ const language = (
+ toolRequest.patch.language ?? (mossLocalGeneration ? audioLanguage : "")
+ ).trim();
saveLastPrompt(ttsDraftKey("prompt", workflow), prompt);
if (savedVoice) {
const voiceController = new AbortController();
@@ -388,7 +391,7 @@ export function useSpeechGeneration({
text,
instructions:
instructionsKind !== null && instructions ? instructions : null,
- language: mossLocalGeneration && language ? language : null,
+ language: language || null,
inputs: { reference: savedVoice },
options: requestOptions,
speed: toolRequest.patch.speed ?? null,
@@ -441,9 +444,7 @@ export function useSpeechGeneration({
...(instructionsKind !== null && instructions
? { audio_instructions: instructions }
: {}),
- ...(mossLocalGeneration && language
- ? { audio_language: language }
- : {}),
+ ...(language ? { audio_language: language } : {}),
...(Object.keys(requestOptions).length > 0
? { audio_options: requestOptions }
: {}),
diff --git a/studio/frontend/src/features/audio/tools/panel-logic.ts b/studio/frontend/src/features/audio/tools/panel-logic.ts
index 8930a5ff7bf..7acca93ce54 100644
--- a/studio/frontend/src/features/audio/tools/panel-logic.ts
+++ b/studio/frontend/src/features/audio/tools/panel-logic.ts
@@ -332,6 +332,26 @@ export const vibeVoiceDialogueLogic: AudioToolPanelLogic<null> = {
toRequest: () => ({}),
};
+// audio.cpp model_specs/supertonic.json; anything else fails its tokenizer.
+export const SUPERTONIC_LANGUAGES: readonly string[] = (
+ "en ko ja ar bg cs da de el es et fi fr hi hr hu id it lt lv nl pl pt ro ru " +
+ "sk sl sv tr uk vi"
+).split(" ");
+
+/** Language code, "" for the model's default (English). */
+export const supertonicLanguageLogic: AudioToolPanelLogic<string> = {
+ id: "supertonic-language",
+ families: ["supertonic"],
+ workflows: ["speak"],
+ title: "Language",
+ claims: [],
+ initial: () => "",
+ toRequest: (value) =>
+ typeof value === "string" && SUPERTONIC_LANGUAGES.includes(value)
+ ? { language: value }
+ : {},
+};
+
export const CLONE_PANEL_LOGIC = [
qwen3TimbreLogic,
indexTts2EmotionLogic,
diff --git a/studio/frontend/src/features/audio/tools/speak-panels.tsx b/studio/frontend/src/features/audio/tools/speak-panels.tsx
index 29de52b5b17..e8231756191 100644
--- a/studio/frontend/src/features/audio/tools/speak-panels.tsx
+++ b/studio/frontend/src/features/audio/tools/speak-panels.tsx
@@ -1,14 +1,23 @@
// SPDX-License-Identifier: AGPL-3.0-only
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+import {
+ Select,
+ SelectContent,
+ SelectItem,
+ SelectTrigger,
+ SelectValue,
+} from "@/components/ui/select";
import { PillTabs } from "@/features/model-picker/components/model-selector/pill-tabs";
import type { AudioOptionSpec } from "../audio-options";
import { AudioOptionFields } from "../audio-options-fields";
import { formatVibeVoiceScript } from "../clone-policy";
import { VoicePicker } from "../components/voice-picker";
import {
+ SUPERTONIC_LANGUAGES,
type SpeakVoiceValue,
speakVoiceLogic,
+ supertonicLanguageLogic,
usesSavedVoice,
vibeVoiceDialogueLogic,
} from "./panel-logic";
@@ -107,7 +116,48 @@ const vibeVoiceDialoguePanel: AudioToolPanel<null> = {
},
};
+const LANGUAGE_NAMES = new Intl.DisplayNames(["en"], { type: "language" });
+// Radix Select cannot hold an empty value.
+const DEFAULT_LANGUAGE = "__default__";
+
+const supertonicLanguagePanel: AudioToolPanel<string> = {
+ ...supertonicLanguageLogic,
+ Component: ({ value, onChange, disabled }) => (
+ <div className="grid gap-1.5">
+ <label
+ className="text-ui-13 font-medium text-foreground"
+ htmlFor="supertonic-language"
+ >
+ Language
+ </label>
+ <Select
+ value={value || DEFAULT_LANGUAGE}
+ onValueChange={(next) =>
+ onChange(next === DEFAULT_LANGUAGE ? "" : next)
+ }
+ disabled={disabled}
+ >
+ <SelectTrigger id="supertonic-language" size="sm" className="w-full">
+ <SelectValue />
+ </SelectTrigger>
+ <SelectContent>
+ <SelectItem value={DEFAULT_LANGUAGE}>English (default)</SelectItem>
+ {SUPERTONIC_LANGUAGES.filter((code) => code !== "en").map((code) => (
+ <SelectItem key={code} value={code}>
+ {LANGUAGE_NAMES.of(code) ?? code}
+ </SelectItem>
+ ))}
+ </SelectContent>
+ </Select>
+ <p className="text-ui-11p5 leading-snug text-muted-foreground">
+ The language of the text, so it is pronounced right.
+ </p>
+ </div>
+ ),
+};
+
export const SPEAK_TOOL_PANELS = [
speakVoicePanel,
vibeVoiceDialoguePanel,
+ supertonicLanguagePanel,
] as const;
diff --git a/studio/frontend/tests/audio-clone-panels.test.ts b/studio/frontend/tests/audio-clone-panels.test.ts
index 67530757d64..4387f5729a0 100644
--- a/studio/frontend/tests/audio-clone-panels.test.ts
+++ b/studio/frontend/tests/audio-clone-panels.test.ts
@@ -30,6 +30,7 @@ const {
qwen3TimbreLogic,
speakVoiceLogic,
speakVoiceStored,
+ supertonicLanguageLogic,
vibeVoiceDialogueLogic,
voiceOptionName,
} = await import("../src/features/audio/tools/panel-logic.ts");
@@ -120,6 +121,36 @@ test("the Voice panel shows on Speak for models that clone or declare a voice",
);
});
+test("Supertonic's Language panel shows on Speak and sends only a code it reads", () => {
+ const supertonic = ctx({
+ audioFamily: "supertonic",
+ audioWorkflows: ["speak"],
+ });
+ assert.equal(
+ panelApplies(supertonicLanguageLogic, "speak", supertonic),
+ true,
+ );
+ assert.equal(
+ panelApplies(
+ supertonicLanguageLogic,
+ "speak",
+ ctx({ audioFamily: "kokoro_tts", audioWorkflows: ["speak"] }),
+ ),
+ false,
+ );
+ const request = (stored: unknown) =>
+ collectToolRequest(
+ [{ ...supertonicLanguageLogic, Component: () => null }],
+ stored === undefined ? {} : { "supertonic-language": stored },
+ { text: "Hallo" },
+ supertonic,
+ ).patch;
+ assert.deepEqual(request("de"), { language: "de" });
+ assert.deepEqual(request(""), {});
+ assert.deepEqual(request(undefined), {});
+ assert.deepEqual(request("German"), {});
+});
+
test("collectToolRequest merges every panel's part and returns the first error", () => {
const a = {
id: "a",