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.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
[Feature] Unsloth Studio / Desktop: let the mic send audio straight to any audio-input model (#16976)
Leo Borcherdingc21ee5702893
diff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.pyindex e42a84ea168..370d4e0f8b2 100644--- a/studio/backend/routes/models.py+++ b/studio/backend/routes/models.py@@ -2153,6 +2153,7 @@ async def list_models(current_subject: str = Depends(get_current_subject)):is_vision = llama_backend.is_vision,is_audio = getattr(llama_backend, "_is_audio", False),audio_type = getattr(llama_backend, "_audio_type", None),+ has_audio_input = bool(getattr(llama_backend, "_has_audio_input", False)),)]diff --git a/studio/backend/tests/test_models_list_gguf_audio_input.py b/studio/backend/tests/test_models_list_gguf_audio_input.pynew file mode 100644index 00000000000..034a47ab98b--- /dev/null+++ b/studio/backend/tests/test_models_list_gguf_audio_input.py@@ -0,0 +1,54 @@+# 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 asyncio+import sys+from pathlib import Path++_BACKEND = Path(__file__).resolve().parents[1]+if str(_BACKEND) not in sys.path:+ sys.path.insert(0, str(_BACKEND))++import routes.inference as inf # noqa: E402+import routes.models as models_route # noqa: E402+++class _FakeLlama:+ is_vision = True+ is_loaded = True+ model_identifier = "unsloth/gemma-4-E2B-it-GGUF"+ _audio_type = "audio_vlm"++ def __init__(self, has_audio_input):+ self._has_audio_input = has_audio_input+++class _FakeUnsloth:+ default_models: list = []+ models: dict = {}+++def _gguf_row(monkeypatch, llama):+ monkeypatch.setattr(models_route, "peek_inference_backend", lambda: _FakeUnsloth())+ monkeypatch.setattr(inf, "get_llama_cpp_backend", lambda: llama)+ (row,) = asyncio.run(models_route.list_models(current_subject = "unsloth")).models+ return row+++def test_gguf_with_an_audio_mmproj_lists_audio_input(monkeypatch):+ row = _gguf_row(monkeypatch, _FakeLlama(True))+ assert row.is_gguf is True+ assert row.has_audio_input is True+++def test_gguf_without_audio_lists_no_audio_input(monkeypatch):+ assert _gguf_row(monkeypatch, _FakeLlama(False)).has_audio_input is False+++def test_a_backend_without_the_attribute_lists_no_audio_input(monkeypatch):+ class _Old(_FakeLlama):+ pass++ llama = _Old(False)+ del llama._has_audio_input+ assert _gguf_row(monkeypatch, llama).has_audio_input is Falsediff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsxindex f8b3f183d44..493fe5caa2d 100644--- a/studio/frontend/src/components/assistant-ui/thread.tsx+++ b/studio/frontend/src/components/assistant-ui/thread.tsx@@ -138,7 +138,9 @@ import {richPasteMarkdown,currentDictationEntryMode,isStudioDictationAvailable,+ micRecordsForModel,notifyStudioDictationUnavailable,+ setMicAudioClipTarget,YoutubeTranscriptPrompt,stripSearchImageTokens,ChatComposerFocusableContext,@@ -370,6 +372,7 @@ import {type PromptQueueRunFailedEventDetail,type PromptQueueStopEventDetail,dictationFailed,+ dictationProducedClip,dictationProducedTranscript,readComposerDraft,type PromptQueueUIEntry,@@ -403,9 +406,14 @@ import {updateStoredChatThread,} from "@/features/chat/utils/chat-history-storage";import {+ clipStillAttached,dictationSendBlocked,shouldSubmitDictation,} from "@/features/chat/utils/dictation-send";+import {+ micSendsAudioToModel,+ pickAudioForModel,+} from "@/features/chat/utils/mic-target";import {isRagClientError,listProjectDocuments,@@ -5415,10 +5423,14 @@ const Composer: FC<{const startDictation = useCallback(() => {if (audioUpload.busy || dictationEntryDisabled) return;if (currentDictationEntryMode() === "recording-file") {- audioUpload.openDialog();+ if (micRecordsForModel()) {+ pickAudioForModel((file) => aui.composer().addAttachment(file));+ } else {+ audioUpload.openDialog();+ }return;}- if (!isStudioDictationAvailable()) {+ if (!micRecordsForModel() && !isStudioDictationAvailable()) {notifyStudioDictationUnavailable();return;}@@ -5428,6 +5440,21 @@ const Composer: FC<{notifyStudioDictationUnavailable();}}, [aui, audioUpload, dictationEntryDisabled]);+ useEffect(+ () =>+ setMicAudioClipTarget({+ add: (file) => aui.composer().addAttachment(file),+ attached: () =>+ aui+ .composer()+ .getState()+ .attachments.flatMap((attachment) =>+ attachment.file ? [attachment.file] : [],+ ),+ }),+ // Per chat: the old chat's unregister discards its unfinished clip.+ [aui, composerIdentity],+ );const sendAfterDictation = useCallback(() => {sendAfterDictationRef.current = true;dictationComposerRef.current = composerIdentity;@@ -5540,6 +5567,10 @@ const Composer: FC<{originComposer: dictationComposerRef.current,currentComposer: composerIdentity,producedTranscript: dictationProducedTranscript(),+ producedClip: clipStillAttached(+ dictationProducedClip(),+ aui.composer().getState().attachments,+ ),baseText: dictationBaseTextRef.current,text,});@@ -7527,6 +7558,16 @@ const ComposerRightControls: FC<{const t = useT();const followUpBehavior = useChatPreferencesStore((s) => s.followUpBehavior);const sendShortcut = useChatPreferencesStore((s) => s.sendShortcut);+ const micTarget = useVoiceSettingsStore((s) => s.micTarget);+ const modelHearsAudio = useChatRuntimeStore(+ (s) =>+ !s.modelLoading &&+ s.models.find((model) => model.id === s.params.checkpoint)+ ?.hasAudioInput === true,+ );+ const micLabel = micSendsAudioToModel(micTarget, modelHearsAudio)+ ? t("settings.voice.dictation.micRecordForModel")+ : "Dictate";// A boolean, so typing re-renders this only when a line break comes or goes.const multiline = useAuiState(({ composer }) => composer.text.includes("\n"));const shortcutLabels = composerShortcutLabels(@@ -7656,8 +7697,8 @@ const ComposerRightControls: FC<{</Button>) : (<TooltipIconButton- tooltip="Dictate"- aria-label="Dictate"+ tooltip={micLabel}+ aria-label={micLabel}type="button"variant="ghost"className="size-9 rounded-full text-foreground"diff --git a/studio/frontend/src/features/chat/adapters/dictation-outcome.ts b/studio/frontend/src/features/chat/adapters/dictation-outcome.tsindex cc231dc24da..12c5bce9276 100644--- a/studio/frontend/src/features/chat/adapters/dictation-outcome.ts+++ b/studio/frontend/src/features/chat/adapters/dictation-outcome.ts@@ -4,13 +4,23 @@/** Module state, read by the recording bar after the session is gone: composer text cannot* tell (a saved prompt may be inserted, a partial transcript still lands). */let producedTranscript = false;+let producedClip: File | null = null;let failed = false;export function beginDictationSession(): void {producedTranscript = false;+ producedClip = null;failed = false;}+export function markDictationClip(file: File): void {+ producedClip = file;+}++export function dictationProducedClip(): File | null {+ return producedClip;+}+export function markDictationTranscript(): void {producedTranscript = true;}diff --git a/studio/frontend/src/features/chat/adapters/mic-clip-state.ts b/studio/frontend/src/features/chat/adapters/mic-clip-state.tsnew file mode 100644index 00000000000..1417df1cdf4--- /dev/null+++ b/studio/frontend/src/features/chat/adapters/mic-clip-state.ts@@ -0,0 +1,17 @@+// 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-session owner: a stale session's late cleanup must not release a newer recording's hold.+let owner: object | null = null;++export function holdMicClipRoom(session: object): void {+ owner = session;+}++export function releaseMicClipRoom(session: object): void {+ if (owner === session) owner = null;+}++export function micClipRecording(): boolean {+ return owner !== null;+}diff --git a/studio/frontend/src/features/chat/adapters/pcm-recorder.ts b/studio/frontend/src/features/chat/adapters/pcm-recorder.tsindex 3ff86d1bafd..6d983cd1029 100644--- a/studio/frontend/src/features/chat/adapters/pcm-recorder.ts+++ b/studio/frontend/src/features/chat/adapters/pcm-recorder.ts@@ -106,6 +106,21 @@ function createAudioContext(): AudioContext {export class PcmRecorder implements SegmentRecorder {readonly mimeType = "audio/wav";readonly sampleRate: number;++ static isSupported(): boolean {+ return (+ typeof window !== "undefined" &&+ window.isSecureContext &&+ typeof (+ window.AudioContext ||+ (window as unknown as { webkitAudioContext?: unknown })+ .webkitAudioContext+ ) === "function" &&+ navigator.mediaDevices?.getUserMedia !== undefined+ );+ }+ /** stop() trims the WAV to this many bytes, header included. */+ maxBytes = Number.POSITIVE_INFINITY;private recordingState: RecordingState = "inactive";private readonly context: AudioContext;private readonly source: MediaStreamAudioSourceNode;@@ -143,6 +158,10 @@ export class PcmRecorder implements SegmentRecorder {return this.recordingState;}+ bytesFor(seconds: number): number {+ return WAV_HEADER_BYTES + Math.ceil(seconds * this.sampleRate) * BYTES_PER_SAMPLE;+ }+/** Seconds that fit in `maxBytes`, less a second of slack for the buffer still in flight. */secondsWithin(maxBytes: number): number {const bytesPerSecond = this.sampleRate * BYTES_PER_SAMPLE;@@ -187,11 +206,14 @@ export class PcmRecorder implements SegmentRecorder {stop(): void {if (this.recordingState === "inactive") return;this.recordingState = "inactive";- const samples = new Float32Array(this.frames);+ const fit = Math.floor((this.maxBytes - WAV_HEADER_BYTES) / BYTES_PER_SAMPLE);+ const samples = new Float32Array(Math.max(0, Math.min(this.frames, fit)));let offset = 0;for (const chunk of this.chunks) {- samples.set(chunk, offset);- offset += chunk.length;+ if (offset >= samples.length) break;+ const part = chunk.subarray(0, samples.length - offset);+ samples.set(part, offset);+ offset += part.length;}this.chunks.length = 0;this.frames = 0;@@ -214,7 +236,19 @@ export class PcmRecorder implements SegmentRecorder {for (const entry of listeners) entry.listener(stopEvent);}+ discard(): void {+ if (this.tornDown) return;+ this.recordingState = "inactive";+ this.chunks.length = 0;+ this.frames = 0;+ this.teardown();+ }++ private tornDown = false;+private teardown(): void {+ if (this.tornDown) return;+ this.tornDown = true;this.source.disconnect();this.processor.disconnect();this.sink.disconnect();diff --git a/studio/frontend/src/features/chat/adapters/studio-audio-clip-adapter.ts b/studio/frontend/src/features/chat/adapters/studio-audio-clip-adapter.tsnew file mode 100644index 00000000000..2eac1dfb495--- /dev/null+++ b/studio/frontend/src/features/chat/adapters/studio-audio-clip-adapter.ts@@ -0,0 +1,233 @@+// 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 { useVoiceSettingsStore } from "@/features/settings/stores/voice-settings-store";+import { MAX_AUDIO_SIZE, MAX_AUDIO_SIZE_LABEL } from "@/lib/audio-utils";+import type { DictationAdapter } from "@assistant-ui/react";+import { toast } from "sonner";+import { type MicClipRoom, micClipFilename } from "../utils/mic-target";+import { startDictationLevelMeter } from "./dictation-level";+import {+ beginDictationSession,+ markDictationClip,+ markDictationFailed,+} from "./dictation-outcome";+import { holdMicClipRoom, releaseMicClipRoom } from "./mic-clip-state";+import { PcmRecorder } from "./pcm-recorder";+import {+ type StudioDictationSession,+ isMissingDeviceError,+} from "./studio-web-speech-dictation-adapter";++// WAV, not MediaRecorder webm: the attachment path routes webm as video.+export function listenForAudioClip(+ onClip: (file: File) => Promise<unknown> | unknown,+ room: MicClipRoom = { maxBytes: MAX_AUDIO_SIZE },+): StudioDictationSession {+ beginDictationSession();+ const speechStartCallbacks = new Set<() => void>();+ const speechEndCallbacks = new Set<+ (result: DictationAdapter.Result) => void+ >();+ const endCallbacks = new Set<() => void>();+ let stream: MediaStream | null = null;+ let recorder: PcmRecorder | null = null;+ let stopLevelMeter = () => {};+ let limitTimer: ReturnType<typeof setTimeout> | undefined;+ let ended = false;+ let resolveEnded: () => void = () => {};+ const endedPromise = new Promise<void>((resolve) => {+ resolveEnded = resolve;+ });++ const release = () => {+ releaseMicClipRoom(session);+ clearTimeout(limitTimer);+ stopLevelMeter();+ for (const track of stream?.getTracks() ?? []) track.stop();+ stream = null;+ };+ const finish = (reason: "stopped" | "cancelled" | "error") => {+ if (ended) return;+ ended = true;+ release();+ session.status = { type: "ended", reason };+ for (const callback of speechEndCallbacks) callback({ transcript: "" });+ for (const callback of endCallbacks) callback();+ resolveEnded();+ };++ const session: StudioDictationSession = {+ status: { type: "starting" },+ stop: async () => {+ const active = recorder;+ recorder = null;+ if (ended || !active || active.state === "inactive") {+ finish("stopped");+ return endedPromise;+ }+ stopLevelMeter();+ let delivering = false;+ // PcmRecorder emits its one blob synchronously from stop().+ active.addEventListener(+ "dataavailable",+ (event) => {+ releaseMicClipRoom(session);+ if (event.data.size === 0) return;+ delivering = true;+ const file = new File([event.data], micClipFilename(), {+ type: "audio/wav",+ });+ void Promise.resolve()+ .then(() => onClip(file))+ .then(+ () => markDictationClip(file),+ (error: unknown) => {+ // The attachment adapter already toasts its own refusals.+ console.error("Could not attach the recording:", error);+ markDictationFailed();+ },+ )+ .finally(() => finish("stopped"));+ },+ { once: true },+ );+ active.stop();+ if (!delivering) finish("stopped");+ return endedPromise;+ },+ cancel: () => {+ const active = recorder;+ recorder = null;+ active?.discard();+ finish("cancelled");+ },+ onSpeechStart: (callback) => {+ speechStartCallbacks.add(callback);+ return () => speechStartCallbacks.delete(callback);+ },+ onSpeechEnd: (callback) => {+ speechEndCallbacks.add(callback);+ return () => speechEndCallbacks.delete(callback);+ },+ onSpeech: () => () => {},+ onEnd: (callback: () => void) => {+ endCallbacks.add(callback);+ return () => endCallbacks.delete(callback);+ },+ };++ if ("reason" in room) {+ toast.error(room.reason);+ markDictationFailed();+ queueMicrotask(() => finish("error"));+ return session;+ }+ const maxBytes = room.maxBytes;+ holdMicClipRoom(session);++ void (async () => {+ const { micDeviceId } = useVoiceSettingsStore.getState();+ const audio: MediaTrackConstraints = {+ echoCancellation: true,+ noiseSuppression: true,+ };+ try {+ try {+ stream = await navigator.mediaDevices.getUserMedia({+ audio:+ micDeviceId && micDeviceId !== "default"+ ? { ...audio, deviceId: { exact: micDeviceId } }+ : audio,+ });+ } catch (error) {+ if (micDeviceId === "default" || !isMissingDeviceError(error)) {+ throw error;+ }+ if (ended) return;+ stream = await navigator.mediaDevices.getUserMedia({ audio });+ }+ if (ended) {+ release();+ return;+ }+ recorder = new PcmRecorder(stream);+ // The pre-check assumes 16 kHz; a device-rate fallback needs more room for the same second.+ if (maxBytes < recorder.bytesFor(1)) {+ recorder.discard();+ recorder = null;+ toast.error(+ `Audio files together exceed the ${MAX_AUDIO_SIZE_LABEL} per-message limit`,+ );+ markDictationFailed();+ finish("error");+ return;+ }+ recorder.maxBytes = maxBytes;+ recorder.start();+ limitTimer = setTimeout(() => {+ toast.info(`Recording stopped at the ${MAX_AUDIO_SIZE_LABEL} audio limit.`);+ void session.stop();+ }, recorder.secondsWithin(maxBytes) * 1000);+ stopLevelMeter = startDictationLevelMeter(stream);+ session.status = { type: "running" };+ for (const callback of speechStartCallbacks) callback();+ } catch (error) {+ // Cancelled already: the outcome flags now belong to the next session.+ if (ended) return;+ console.error("Microphone error:", error);+ toast.error(+ isMissingDeviceError(error)+ ? "No microphone was found."+ : "Could not access the microphone.",+ );+ markDictationFailed();+ finish("error");+ }+ })();++ return session;+}++type ModelState = {+ params: { checkpoint: string | null | undefined };+ modelLoading?: boolean;+ loadingModelPick?: unknown;+};+type CheckpointStore = {+ getState: () => ModelState;+ subscribe: (listener: (state: ModelState) => void) => () => void;+};++/** A load starting counts as a change: `checkpoint` only moves once the load succeeds. */+export function discardClipWhenModelChanges(+ session: StudioDictationSession,+ store: CheckpointStore,+): void {+ const startedOn = store.getState().params.checkpoint;+ const unsubscribe = store.subscribe((state) => {+ const switching =+ state.params.checkpoint !== startedOn ||+ state.modelLoading === true ||+ state.loadingModelPick != null;+ if (!switching) return;+ unsubscribe();+ if (session.status.type === "ended") return;+ toast.info("Recording discarded because the chat model changed.");+ session.cancel();+ });+ session.onEnd?.(unsubscribe);+}++export function discardClipOnRelease(+ session: StudioDictationSession,+ onRelease: Set<() => void>,+): void {+ const release = () => {+ if (session.status.type === "ended") return;+ toast.info("Recording discarded because the chat changed.");+ session.cancel();+ };+ onRelease.add(release);+ session.onEnd?.(() => onRelease.delete(release));+}diff --git a/studio/frontend/src/features/chat/adapters/studio-dictation-adapter.tsx b/studio/frontend/src/features/chat/adapters/studio-dictation-adapter.tsxindex 18142100be1..822d39b8b3f 100644--- a/studio/frontend/src/features/chat/adapters/studio-dictation-adapter.tsx+++ b/studio/frontend/src/features/chat/adapters/studio-dictation-adapter.tsx@@ -9,7 +9,10 @@ import {} from "@/features/settings/stores/voice-settings-store";import { toast } from "@/lib/toast";import type { DictationAdapter } from "@assistant-ui/react";+import { useChatRuntimeStore } from "../stores/chat-runtime-store";import { useExternalProvidersStore } from "../stores/external-providers-store";+import { maxAudioFilesFor } from "@/lib/audio-utils";+import { micClipRoom, micSendsAudioToModel } from "../utils/mic-target";import {currentDictationEntryMode,insecureDictationGuidance,@@ -20,6 +23,12 @@ import {sttEngineStatusFor,sttQuantDownloaded,} from "./studio-model-dictation-adapter";+import { PcmRecorder } from "./pcm-recorder";+import {+ discardClipOnRelease,+ discardClipWhenModelChanges,+ listenForAudioClip,+} from "./studio-audio-clip-adapter";import {type StudioDictationSession,StudioWebSpeechDictationAdapter,@@ -35,6 +44,33 @@ export function cancelActiveStudioDictation(): void {session?.cancel();}+type AudioClipTarget = {+ add: (file: File) => Promise<unknown> | unknown;+ attached: () => readonly { name: string; type: string; size: number }[];+};+type RegisteredClipTarget = AudioClipTarget & { onRelease: Set<() => void> };+// Compare registers none, so it keeps transcribing.+let audioClipTarget: RegisteredClipTarget | null = null;++export function setMicAudioClipTarget(target: AudioClipTarget): () => void {+ const entry: RegisteredClipTarget = { ...target, onRelease: new Set() };+ audioClipTarget = entry;+ return () => {+ if (audioClipTarget === entry) audioClipTarget = null;+ for (const listener of [...entry.onRelease]) listener();+ };+}++export function micRecordsForModel(chatId?: string | null): boolean {+ if (chatId === null || !audioClipTarget) return false;+ const { micTarget } = useVoiceSettingsStore.getState();+ const { params, models, modelLoading } = useChatRuntimeStore.getState();+ const activeModel = modelLoading+ ? undefined+ : models.find((model) => model.id === params.checkpoint);+ return micSendsAudioToModel(micTarget, activeModel?.hasAudioInput === true);+}+/** Local and custom transcription record through the media recorder. */function usesRecordedAudio(dictationEngine: DictationEngine): boolean {return dictationEngine !== "browser";@@ -89,6 +125,26 @@ export class StudioDictationAdapter implements DictationAdapter {}private createSession(): StudioDictationSession {+ const clipTarget = audioClipTarget;+ if (clipTarget && micRecordsForModel(this.chatId)) {+ if (!PcmRecorder.isSupported()) {+ throw new Error("Voice recording is not supported in this browser.");+ }+ const { params, models, modelLoading, pendingAudioBase64 } =+ useChatRuntimeStore.getState();+ const activeModel = modelLoading+ ? undefined+ : models.find((model) => model.id === params.checkpoint);+ const room = micClipRoom(+ clipTarget.attached(),+ maxAudioFilesFor(activeModel),+ !!pendingAudioBase64,+ );+ const session = listenForAudioClip(clipTarget.add, room);+ discardClipWhenModelChanges(session, useChatRuntimeStore);+ discardClipOnRelease(session, clipTarget.onRelease);+ return session;+ }const { dictationEngine } = useVoiceSettingsStore.getState();if (usesRecordedAudio(dictationEngine)) {if (dictationEngine === "custom" && !customSttConfigured()) {diff --git a/studio/frontend/src/features/chat/audio-attachment-adapter.ts b/studio/frontend/src/features/chat/audio-attachment-adapter.tsindex eb7c0e602b2..e674c801cd9 100644--- a/studio/frontend/src/features/chat/audio-attachment-adapter.ts+++ b/studio/frontend/src/features/chat/audio-attachment-adapter.ts@@ -14,6 +14,7 @@ import type {PendingAttachment,} from "@assistant-ui/react";import { toast } from "sonner";+import { micClipRecording } from "./adapters/mic-clip-state";import { externalModelLabel } from "./lib/external-model-label";import { useChatRuntimeStore } from "./stores/chat-runtime-store";@@ -60,6 +61,11 @@ export class AudioAttachmentAdapter implements AttachmentAdapter {toast.error(stagedReason, { id: AUDIO_ADD_TOAST_ID });throw new Error(stagedReason);}+ if (micClipRecording()) {+ const recordingReason = "Stop the recording before attaching more audio.";+ toast.error(recordingReason, { id: AUDIO_ADD_TOAST_ID });+ throw new Error(recordingReason);+ }let totalSize = 0;for (const size of this.attachmentSizes.values()) totalSize += size;const addReason = getAudioAddError(diff --git a/studio/frontend/src/features/chat/index.ts b/studio/frontend/src/features/chat/index.tsindex 9b3ce5548cd..fd7868f4655 100644--- a/studio/frontend/src/features/chat/index.ts+++ b/studio/frontend/src/features/chat/index.ts@@ -611,13 +611,16 @@ export {export { subscribeDictationLevel } from "./adapters/dictation-level";export {dictationFailed,+ dictationProducedClip,dictationProducedTranscript,} from "./adapters/dictation-outcome";export {StudioDictationAdapter,cancelActiveStudioDictation,isStudioDictationAvailable,+ micRecordsForModel,notifyStudioDictationUnavailable,+ setMicAudioClipTarget,} from "./adapters/studio-dictation-adapter";export {StudioModelDictationAdapter,diff --git a/studio/frontend/src/features/chat/utils/dictation-send.ts b/studio/frontend/src/features/chat/utils/dictation-send.tsindex d74fec7decd..9e1c6ffc448 100644--- a/studio/frontend/src/features/chat/utils/dictation-send.ts+++ b/studio/frontend/src/features/chat/utils/dictation-send.ts@@ -29,15 +29,24 @@ export function dictationSendBlocked(state: {);}-/** Refuse on a thread switch, a text change not from speech, or silence. */+/** Refuse on a thread switch, a text change not from speech, or silence; an attached clip always sends. */export function shouldSubmitDictation(input: {originComposer: string;currentComposer: string;producedTranscript: boolean;+ producedClip?: boolean;baseText: string;text: string;}): boolean {if (input.originComposer !== input.currentComposer) return false;+ if (input.producedClip) return true;if (!input.producedTranscript) return false;return dictationProducedText(input.baseText, input.text);}++export function clipStillAttached(+ clip: File | null,+ attachments: readonly { file?: File }[],+): boolean {+ return clip !== null && attachments.some((attachment) => attachment.file === clip);+}diff --git a/studio/frontend/src/features/chat/utils/mic-target.ts b/studio/frontend/src/features/chat/utils/mic-target.tsnew file mode 100644index 00000000000..86793f43eec--- /dev/null+++ b/studio/frontend/src/features/chat/utils/mic-target.ts@@ -0,0 +1,71 @@+// 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 type { MicTarget } from "@/features/settings/stores/voice-settings-store";+import {+ AUDIO_ATTACHMENT_ACCEPT,+ MAX_AUDIO_SIZE,+ getAudioAddError,+ isAudioAttachmentFile,+} from "../../../lib/audio-utils.ts";++export function micSendsAudioToModel(+ micTarget: MicTarget,+ modelHearsAudio: boolean,+): boolean {+ return micTarget === "model" && modelHearsAudio;+}++export function micClipFilename(now: Date = new Date()): string {+ const stamp = now.toISOString().slice(0, 19).replaceAll(":", "-");+ return `recording-${stamp}.wav`;+}++export const MIN_MIC_CLIP_BYTES = 44 + 16_000 * 2;++export type MicClipRoom = { maxBytes: number } | { reason: string };++export function micClipRoom(+ attached: readonly { name: string; type: string; size: number }[],+ maxFiles: number,+ stagedClip: boolean,+): MicClipRoom {+ if (stagedClip) {+ return {+ reason: "Send or remove the staged audio clip before attaching more audio.",+ };+ }+ const audio = attached.filter(isAudioAttachmentFile);+ const totalSize = audio.reduce((sum, file) => sum + file.size, 0);+ const reason = getAudioAddError(+ audio.length,+ totalSize,+ MIN_MIC_CLIP_BYTES,+ maxFiles,+ );+ return reason ? { reason } : { maxBytes: MAX_AUDIO_SIZE - totalSize };+}++export function pickAudioForModel(+ add: (file: File) => Promise<unknown>,+ doc: Document = document,+): void {+ const input = doc.createElement("input");+ input.type = "file";+ input.accept = AUDIO_ATTACHMENT_ACCEPT;+ input.multiple = true;+ input.hidden = true;+ const done = () => input.remove();+ input.addEventListener(+ "change",+ () => {+ // The attachment adapter toasts its own refusals.+ for (const file of Array.from(input.files ?? [])) void add(file).catch(() => {});+ done();+ },+ { once: true },+ );+ input.addEventListener("cancel", done, { once: true });+ doc.body.append(input);+ input.click();+}diff --git a/studio/frontend/src/features/settings/stores/voice-settings-store.ts b/studio/frontend/src/features/settings/stores/voice-settings-store.tsindex 967b84459f0..07ecf5fd3f9 100644--- a/studio/frontend/src/features/settings/stores/voice-settings-store.ts+++ b/studio/frontend/src/features/settings/stores/voice-settings-store.ts@@ -82,6 +82,9 @@ export function isSttModelLanguageCompatible(export type DictationEngine = "browser" | "model" | "custom";+/** "model" attaches a clip for an audio-input model, else falls back to "asr". */+export type MicTarget = "asr" | "model";+/** Mirrors the backend `SttLoadRequest.device`; "gpu" is not offered since it adds nothing. */export type SttDevice = "auto" | "cpu";@@ -107,6 +110,9 @@ export interface VoiceSettingsState {dictationEngine: DictationEngine;setDictationEngine: (value: DictationEngine) => void;+ micTarget: MicTarget;+ setMicTarget: (value: MicTarget) => void;+/** STT model used when dictationEngine is "model". */sttModel: SttModel;setSttModel: (value: SttModel) => void;@@ -226,6 +232,9 @@ export const useVoiceSettingsStore = create<VoiceSettingsState>()(dictationEngine: isTauri ? "model" : "browser",setDictationEngine: (dictationEngine) => set({ dictationEngine }),+ micTarget: "asr",+ setMicTarget: (micTarget) => set({ micTarget }),+sttModel: DEFAULT_STT_MODEL,setSttModel: (value) =>set((state) => {@@ -397,6 +406,7 @@ export const useVoiceSettingsStore = create<VoiceSettingsState>()(...current,micDeviceId: asString(saved?.micDeviceId, "default"),dictationEngine,+ micTarget: saved?.micTarget === "model" ? "model" : "asr",sttModel,sttGgufVariant:sttModel === savedSttModeldiff --git a/studio/frontend/src/features/settings/tabs/voice-tab.tsx b/studio/frontend/src/features/settings/tabs/voice-tab.tsxindex fd6998cea17..8b7360dc286 100644--- a/studio/frontend/src/features/settings/tabs/voice-tab.tsx+++ b/studio/frontend/src/features/settings/tabs/voice-tab.tsx@@ -43,6 +43,7 @@ import {sttEngineFor,sttEngineStatusFor,unloadSttModel,+ useChatRuntimeStore,useExternalProvidersStore,validateSttModel,} from "@/features/chat";@@ -487,6 +488,14 @@ export function VoiceTab() {const micDeviceId = useVoiceSettingsStore((s) => s.micDeviceId);const setMicDeviceId = useVoiceSettingsStore((s) => s.setMicDeviceId);const dictationEngine = useVoiceSettingsStore((s) => s.dictationEngine);+ const micTarget = useVoiceSettingsStore((s) => s.micTarget);+ const setMicTarget = useVoiceSettingsStore((s) => s.setMicTarget);+ const modelHearsAudio = useChatRuntimeStore(+ (s) =>+ !s.modelLoading &&+ s.models.find((model) => model.id === s.params.checkpoint)+ ?.hasAudioInput === true,+ );const setDictationEngine = useVoiceSettingsStore((s) => s.setDictationEngine);const sttModel = useVoiceSettingsStore((s) => s.sttModel);const setSttModel = useVoiceSettingsStore((s) => s.setSttModel);@@ -1171,6 +1180,44 @@ export function VoiceTab() {</header><SettingsSection title={t("settings.voice.dictation.sectionTitle")}>+ <SettingsRow+ label={t("settings.voice.dictation.micTargetLabel")}+ description={+ !modelHearsAudio+ ? t("settings.voice.dictation.micTargetModelUnavailable")+ : micTarget === "model"+ ? t("settings.voice.dictation.micTargetModelDescription")+ : t("settings.voice.dictation.micTargetAsrDescription")+ }+ >+ <Select+ value={micTarget}+ onValueChange={(value) =>+ setMicTarget(value === "model" ? "model" : "asr")+ }+ >+ <SelectTrigger+ data-testid="mic-target-trigger"+ aria-label={t("settings.voice.dictation.micTargetLabel")}+ className="w-56"+ size="sm"+ >+ <SelectValue />+ </SelectTrigger>+ <SelectContent>+ <SelectItem value="asr">+ {t("settings.voice.dictation.micTargetAsr")}+ </SelectItem>+ <SelectItem+ value="model"+ data-testid="mic-target-model"+ disabled={!modelHearsAudio && micTarget !== "model"}+ >+ {t("settings.voice.dictation.micTargetModel")}+ </SelectItem>+ </SelectContent>+ </Select>+ </SettingsRow><SettingsRowlabel={t("settings.voice.dictation.engineLabel")}description={diff --git a/studio/frontend/src/i18n/locales/ar.ts b/studio/frontend/src/i18n/locales/ar.tsindex 7c0a42cd4c2..8afcc3b8cba 100644--- a/studio/frontend/src/i18n/locales/ar.ts+++ b/studio/frontend/src/i18n/locales/ar.ts@@ -1661,6 +1661,13 @@ export const ar = {audioUploadCancel: "إلغاء التفريغ النصي",audioUploadTranscribing: "جارٍ التفريغ النصي للصوت…",sectionTitle: "الإملاء",+ micTargetLabel: "زر الميكروفون",+ micTargetAsr: "النسخ باستخدام ASR",+ micTargetModel: "إرسال الصوت إلى النموذج",+ micTargetAsrDescription: "يكتب زر الميكروفون ما تقوله في مربع الرسالة.",+ micTargetModelDescription: "عند تحميل نموذج يقبل إدخال الصوت، يسجّل الميكروفون مقطعًا ويرفقه برسالتك. تحصل النماذج الأخرى على الإملاء كالمعتاد.",+ micTargetModelUnavailable: "حمّل نموذجًا يقبل إدخال الصوت (مثل LFM3-Omni) لترسل إليه صوتك.",+ micRecordForModel: "تسجيل صوت للنموذج",engineLabel: "محرك الإملاء",engineBrowser: "المتصفح",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/cs.ts b/studio/frontend/src/i18n/locales/cs.tsindex 90cd157a369..ccc37e96c20 100644--- a/studio/frontend/src/i18n/locales/cs.ts+++ b/studio/frontend/src/i18n/locales/cs.ts@@ -1834,6 +1834,13 @@ export const cs = {audioUploadCancel: "Zrušit přepis",audioUploadTranscribing: "Přepisování zvuku…",sectionTitle: "Diktování",+ micTargetLabel: "Tlačítko mikrofonu",+ micTargetAsr: "Přepsat pomocí ASR",+ micTargetModel: "Odeslat zvuk modelu",+ micTargetAsrDescription: "Tlačítko mikrofonu zapíše, co řeknete, do pole zprávy.",+ micTargetModelDescription: "Když je načten model se zvukovým vstupem, mikrofon nahraje klip a připojí ho ke zprávě. Ostatní modely dál používají diktování.",+ micTargetModelUnavailable: "Načtěte model, který přijímá zvukový vstup (například LFM3-Omni), abyste mu mohli poslat svůj hlas.",+ micRecordForModel: "Nahrát zvuk pro model",engineLabel: "Modul diktování",engineBrowser: "Prohlížeč",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/de.ts b/studio/frontend/src/i18n/locales/de.tsindex 9f1cb20d1e3..10707ba53f0 100644--- a/studio/frontend/src/i18n/locales/de.ts+++ b/studio/frontend/src/i18n/locales/de.ts@@ -1665,6 +1665,13 @@ export const de = {audioUploadCancel: "Transkription abbrechen",audioUploadTranscribing: "Audio wird transkribiert…",sectionTitle: "Diktat",+ micTargetLabel: "Mikrofontaste",+ micTargetAsr: "Mit ASR transkribieren",+ micTargetModel: "Audio an das Modell senden",+ micTargetAsrDescription: "Die Mikrofontaste schreibt, was du sagst, in das Nachrichtenfeld.",+ micTargetModelDescription: "Ist ein Modell mit Audioeingabe geladen, nimmt das Mikrofon einen Clip auf und hängt ihn an deine Nachricht an. Andere Modelle erhalten weiterhin Diktat.",+ micTargetModelUnavailable: "Lade ein Modell mit Audioeingabe (zum Beispiel LFM3-Omni), um ihm deine Stimme zu senden.",+ micRecordForModel: "Audio für das Modell aufnehmen",engineLabel: "Diktat-Engine",engineBrowser: "Browser",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.tsindex 61fc02c810c..5c487802c00 100644--- a/studio/frontend/src/i18n/locales/en.ts+++ b/studio/frontend/src/i18n/locales/en.ts@@ -1670,6 +1670,16 @@ export const en = {audioUploadCancel: "Cancel transcription",audioUploadTranscribing: "Transcribing audio…",sectionTitle: "Dictation",+ micTargetLabel: "Mic button",+ micTargetAsr: "Transcribe with ASR",+ micTargetModel: "Send audio to the model",+ micTargetAsrDescription:+ "The mic button types what you say into the message box.",+ micTargetModelDescription:+ "With an audio-input model loaded, the mic records a clip and attaches it to your message. Other models still get dictation.",+ micTargetModelUnavailable:+ "Load a model that accepts audio input (for example LFM3-Omni) to send it your voice.",+ micRecordForModel: "Record audio for the model",engineLabel: "Dictation engine",engineBrowser: "Browser",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/es-419.ts b/studio/frontend/src/i18n/locales/es-419.tsindex b9a82a4284e..d9415a8f17d 100644--- a/studio/frontend/src/i18n/locales/es-419.ts+++ b/studio/frontend/src/i18n/locales/es-419.ts@@ -1845,6 +1845,13 @@ export const es419 = {audioUploadCancel: "Cancelar transcripción",audioUploadTranscribing: "Transcribiendo audio…",sectionTitle: "Dictado",+ micTargetLabel: "Botón del micrófono",+ micTargetAsr: "Transcribir con ASR",+ micTargetModel: "Enviar audio al modelo",+ micTargetAsrDescription: "El botón del micrófono escribe lo que dices en el cuadro de mensaje.",+ micTargetModelDescription: "Con un modelo con entrada de audio cargado, el micrófono graba un clip y lo adjunta a tu mensaje. Los demás modelos siguen usando dictado.",+ micTargetModelUnavailable: "Carga un modelo que acepte entrada de audio (por ejemplo, LFM3-Omni) para enviarle tu voz.",+ micRecordForModel: "Grabar audio para el modelo",engineLabel: "Motor de dictado",engineBrowser: "Navegador",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/es.ts b/studio/frontend/src/i18n/locales/es.tsindex 24b3cb76632..de5ead339ba 100644--- a/studio/frontend/src/i18n/locales/es.ts+++ b/studio/frontend/src/i18n/locales/es.ts@@ -1664,6 +1664,13 @@ export const es = {audioUploadCancel: "Cancelar transcripción",audioUploadTranscribing: "Transcribiendo audio…",sectionTitle: "Dictado",+ micTargetLabel: "Botón del micrófono",+ micTargetAsr: "Transcribir con ASR",+ micTargetModel: "Enviar audio al modelo",+ micTargetAsrDescription: "El botón del micrófono escribe lo que dices en el cuadro de mensaje.",+ micTargetModelDescription: "Con un modelo con entrada de audio cargado, el micrófono graba un clip y lo adjunta a tu mensaje. Los demás modelos siguen usando dictado.",+ micTargetModelUnavailable: "Carga un modelo que acepte entrada de audio (por ejemplo, LFM3-Omni) para enviarle tu voz.",+ micRecordForModel: "Grabar audio para el modelo",engineLabel: "Motor de dictado",engineBrowser: "Navegador",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/fa.ts b/studio/frontend/src/i18n/locales/fa.tsindex 86a286a6895..2558f765d28 100644--- a/studio/frontend/src/i18n/locales/fa.ts+++ b/studio/frontend/src/i18n/locales/fa.ts@@ -1819,6 +1819,13 @@ export const fa = {audioUploadCancel: "لغو رونویسی",audioUploadTranscribing: "در حال رونویسی صدا…",sectionTitle: "دیکته",+ micTargetLabel: "دکمه میکروفون",+ micTargetAsr: "رونویسی با ASR",+ micTargetModel: "ارسال صدا به مدل",+ micTargetAsrDescription: "دکمه میکروفون آنچه میگویید را در کادر پیام تایپ میکند.",+ micTargetModelDescription: "وقتی مدلی با ورودی صوتی بارگذاری شده باشد، میکروفون یک کلیپ ضبط کرده و به پیام شما پیوست میکند. مدلهای دیگر همچنان دیکته دریافت میکنند.",+ micTargetModelUnavailable: "برای ارسال صدای خود، مدلی را بارگذاری کنید که ورودی صوتی میپذیرد (برای مثال LFM3-Omni).",+ micRecordForModel: "ضبط صدا برای مدل",engineLabel: "موتور دیکته",engineBrowser: "مرورگر",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/fr.ts b/studio/frontend/src/i18n/locales/fr.tsindex 1131a626847..9dccbcb2c17 100644--- a/studio/frontend/src/i18n/locales/fr.ts+++ b/studio/frontend/src/i18n/locales/fr.ts@@ -1665,6 +1665,13 @@ export const fr = {audioUploadCancel: "Annuler la transcription",audioUploadTranscribing: "Transcription audio en cours…",sectionTitle: "Dictée",+ micTargetLabel: "Bouton du micro",+ micTargetAsr: "Transcrire avec l'ASR",+ micTargetModel: "Envoyer l'audio au modèle",+ micTargetAsrDescription: "Le bouton du micro écrit ce que vous dites dans la zone de message.",+ micTargetModelDescription: "Avec un modèle à entrée audio chargé, le micro enregistre un extrait et le joint à votre message. Les autres modèles utilisent toujours la dictée.",+ micTargetModelUnavailable: "Chargez un modèle qui accepte l'entrée audio (par exemple LFM3-Omni) pour lui envoyer votre voix.",+ micRecordForModel: "Enregistrer l'audio pour le modèle",engineLabel: "Moteur de dictée",engineBrowser: "Navigateur",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/he.ts b/studio/frontend/src/i18n/locales/he.tsindex 84675c1f3b3..12ada630ea9 100644--- a/studio/frontend/src/i18n/locales/he.ts+++ b/studio/frontend/src/i18n/locales/he.ts@@ -1730,6 +1730,13 @@ export const he = {audioUploadCancel: "ביטול תמלול",audioUploadTranscribing: "מתמלל אודיו…",sectionTitle: "הכתבה",+ micTargetLabel: "כפתור המיקרופון",+ micTargetAsr: "תמלול באמצעות ASR",+ micTargetModel: "שליחת שמע למודל",+ micTargetAsrDescription: "כפתור המיקרופון מקליד את מה שאתם אומרים בתיבת ההודעה.",+ micTargetModelDescription: "כאשר טעון מודל שמקבל קלט שמע, המיקרופון מקליט קטע ומצרף אותו להודעה. מודלים אחרים ממשיכים לקבל הכתבה.",+ micTargetModelUnavailable: "טענו מודל שמקבל קלט שמע (לדוגמה LFM3-Omni) כדי לשלוח לו את הקול שלכם.",+ micRecordForModel: "הקלטת שמע למודל",engineLabel: "מנוע הכתבה",engineBrowser: "דפדפן",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/hi.ts b/studio/frontend/src/i18n/locales/hi.tsindex b70510fed7e..e57781da875 100644--- a/studio/frontend/src/i18n/locales/hi.ts+++ b/studio/frontend/src/i18n/locales/hi.ts@@ -1665,6 +1665,13 @@ export const hi = {audioUploadCancel: "ट्रांसक्रिप्शन रद्द करें",audioUploadTranscribing: "ऑडियो को टेक्स्ट में बदला जा रहा है…",sectionTitle: "डिक्टेशन",+ micTargetLabel: "माइक बटन",+ micTargetAsr: "ASR से ट्रांसक्राइब करें",+ micTargetModel: "मॉडल को ऑडियो भेजें",+ micTargetAsrDescription: "माइक बटन आप जो बोलते हैं उसे संदेश बॉक्स में टाइप करता है।",+ micTargetModelDescription: "ऑडियो इनपुट वाला मॉडल लोड होने पर माइक एक क्लिप रिकॉर्ड करके आपके संदेश में जोड़ता है। अन्य मॉडलों के लिए डिक्टेशन ही चलता है।",+ micTargetModelUnavailable: "अपनी आवाज़ भेजने के लिए ऑडियो इनपुट स्वीकार करने वाला मॉडल लोड करें (जैसे LFM3-Omni)।",+ micRecordForModel: "मॉडल के लिए ऑडियो रिकॉर्ड करें",engineLabel: "डिक्टेशन इंजन",engineBrowser: "ब्राउज़र",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/id.ts b/studio/frontend/src/i18n/locales/id.tsindex d3ec1b6718c..6a0b0982369 100644--- a/studio/frontend/src/i18n/locales/id.ts+++ b/studio/frontend/src/i18n/locales/id.ts@@ -1830,6 +1830,13 @@ export const id = {audioUploadCancel: "Batalkan transkripsi",audioUploadTranscribing: "Mentranskripsikan audio…",sectionTitle: "Dikte",+ micTargetLabel: "Tombol mikrofon",+ micTargetAsr: "Transkripsi dengan ASR",+ micTargetModel: "Kirim audio ke model",+ micTargetAsrDescription: "Tombol mikrofon mengetik ucapan Anda ke kotak pesan.",+ micTargetModelDescription: "Saat model dengan input audio dimuat, mikrofon merekam klip dan melampirkannya ke pesan Anda. Model lain tetap menggunakan dikte.",+ micTargetModelUnavailable: "Muat model yang menerima input audio (misalnya LFM3-Omni) untuk mengirim suara Anda.",+ micRecordForModel: "Rekam audio untuk model",engineLabel: "Engine dikte",engineBrowser: "Browser",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/it.ts b/studio/frontend/src/i18n/locales/it.tsindex 35029ef3165..9863501eab2 100644--- a/studio/frontend/src/i18n/locales/it.ts+++ b/studio/frontend/src/i18n/locales/it.ts@@ -1563,6 +1563,13 @@ export const it = {audioUploadCancel: "Annulla trascrizione",audioUploadTranscribing: "Trascrizione audio in corso…",sectionTitle: "Dettatura",+ micTargetLabel: "Pulsante del microfono",+ micTargetAsr: "Trascrivi con ASR",+ micTargetModel: "Invia l'audio al modello",+ micTargetAsrDescription: "Il pulsante del microfono scrive ciò che dici nella casella del messaggio.",+ micTargetModelDescription: "Con un modello con input audio caricato, il microfono registra una clip e la allega al messaggio. Gli altri modelli continuano a usare la dettatura.",+ micTargetModelUnavailable: "Carica un modello che accetta input audio (ad esempio LFM3-Omni) per inviargli la tua voce.",+ micRecordForModel: "Registra audio per il modello",engineLabel: "Motore di dettatura",engineBrowser: "Browser",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/ja.ts b/studio/frontend/src/i18n/locales/ja.tsindex 004c71576ad..67bb2e43f47 100644--- a/studio/frontend/src/i18n/locales/ja.ts+++ b/studio/frontend/src/i18n/locales/ja.ts@@ -1663,6 +1663,13 @@ export const ja = {audioUploadCancel: "文字起こしをキャンセル",audioUploadTranscribing: "音声を文字起こし中…",sectionTitle: "音声入力",+ micTargetLabel: "マイクボタン",+ micTargetAsr: "ASR で文字起こし",+ micTargetModel: "音声をモデルに送信",+ micTargetAsrDescription: "マイクボタンで話した内容をメッセージ欄に入力します。",+ micTargetModelDescription: "音声入力対応のモデルを読み込んでいると、マイクでクリップを録音してメッセージに添付します。その他のモデルでは従来どおり音声入力になります。",+ micTargetModelUnavailable: "音声を送るには、音声入力に対応したモデル(例: LFM3-Omni)を読み込んでください。",+ micRecordForModel: "モデル用に音声を録音",engineLabel: "音声入力エンジン",engineBrowser: "ブラウザー",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/ko.ts b/studio/frontend/src/i18n/locales/ko.tsindex 382fd19dbec..1cb17cf0338 100644--- a/studio/frontend/src/i18n/locales/ko.ts+++ b/studio/frontend/src/i18n/locales/ko.ts@@ -1661,6 +1661,13 @@ export const ko = {audioUploadCancel: "전사 취소",audioUploadTranscribing: "오디오 전사 중…",sectionTitle: "받아쓰기",+ micTargetLabel: "마이크 버튼",+ micTargetAsr: "ASR로 받아쓰기",+ micTargetModel: "모델에 오디오 보내기",+ micTargetAsrDescription: "마이크 버튼이 말한 내용을 메시지 상자에 입력합니다.",+ micTargetModelDescription: "오디오 입력을 지원하는 모델이 로드되어 있으면 마이크가 클립을 녹음해 메시지에 첨부합니다. 다른 모델은 계속 받아쓰기를 사용합니다.",+ micTargetModelUnavailable: "음성을 보내려면 오디오 입력을 지원하는 모델(예: LFM3-Omni)을 로드하세요.",+ micRecordForModel: "모델용 오디오 녹음",engineLabel: "받아쓰기 엔진",engineBrowser: "브라우저",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/nl.ts b/studio/frontend/src/i18n/locales/nl.tsindex 5d7a7f24c83..b2a488aea7f 100644--- a/studio/frontend/src/i18n/locales/nl.ts+++ b/studio/frontend/src/i18n/locales/nl.ts@@ -1851,6 +1851,13 @@ export const nl = {audioUploadCancel: "Transcriptie annuleren",audioUploadTranscribing: "Audio transcriberen…",sectionTitle: "Dicteren",+ micTargetLabel: "Microfoonknop",+ micTargetAsr: "Transcriberen met ASR",+ micTargetModel: "Audio naar het model sturen",+ micTargetAsrDescription: "De microfoonknop typt wat je zegt in het berichtvak.",+ micTargetModelDescription: "Met een model met audio-invoer geladen neemt de microfoon een clip op en voegt die toe aan je bericht. Andere modellen krijgen nog steeds dicteren.",+ micTargetModelUnavailable: "Laad een model dat audio-invoer accepteert (bijvoorbeeld LFM3-Omni) om het je stem te sturen.",+ micRecordForModel: "Audio opnemen voor het model",engineLabel: "Dicteerengine",engineBrowser: "Browser",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/pl.ts b/studio/frontend/src/i18n/locales/pl.tsindex 031d5bdac25..bcc12371fb6 100644--- a/studio/frontend/src/i18n/locales/pl.ts+++ b/studio/frontend/src/i18n/locales/pl.ts@@ -1845,6 +1845,13 @@ export const pl = {audioUploadCancel: "Anuluj transkrypcję",audioUploadTranscribing: "Transkrybowanie dźwięku…",sectionTitle: "Dyktowanie",+ micTargetLabel: "Przycisk mikrofonu",+ micTargetAsr: "Transkrybuj za pomocą ASR",+ micTargetModel: "Wyślij dźwięk do modelu",+ micTargetAsrDescription: "Przycisk mikrofonu wpisuje to, co mówisz, w polu wiadomości.",+ micTargetModelDescription: "Gdy załadowany jest model z wejściem audio, mikrofon nagrywa klip i dołącza go do wiadomości. Pozostałe modele nadal korzystają z dyktowania.",+ micTargetModelUnavailable: "Załaduj model obsługujący wejście audio (na przykład LFM3-Omni), aby wysłać mu swój głos.",+ micRecordForModel: "Nagraj dźwięk dla modelu",engineLabel: "Silnik dyktowania",engineBrowser: "Przeglądarka",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/pt-br.ts b/studio/frontend/src/i18n/locales/pt-br.tsindex a5510828682..6a7f48de78e 100644--- a/studio/frontend/src/i18n/locales/pt-br.ts+++ b/studio/frontend/src/i18n/locales/pt-br.ts@@ -1663,6 +1663,13 @@ export const ptBR = {audioUploadCancel: "Cancelar transcrição",audioUploadTranscribing: "Transcrevendo áudio…",sectionTitle: "Ditado",+ micTargetLabel: "Botão do microfone",+ micTargetAsr: "Transcrever com ASR",+ micTargetModel: "Enviar áudio ao modelo",+ micTargetAsrDescription: "O botão do microfone digita o que você fala na caixa de mensagem.",+ micTargetModelDescription: "Com um modelo com entrada de áudio carregado, o microfone grava um clipe e o anexa à sua mensagem. Os outros modelos continuam usando ditado.",+ micTargetModelUnavailable: "Carregue um modelo que aceite entrada de áudio (por exemplo, LFM3-Omni) para enviar sua voz a ele.",+ micRecordForModel: "Gravar áudio para o modelo",engineLabel: "Mecanismo de ditado",engineBrowser: "Navegador",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/ru.ts b/studio/frontend/src/i18n/locales/ru.tsindex a42f1ea092d..991adbc3617 100644--- a/studio/frontend/src/i18n/locales/ru.ts+++ b/studio/frontend/src/i18n/locales/ru.ts@@ -1663,6 +1663,13 @@ export const ru = {audioUploadCancel: "Отменить расшифровку",audioUploadTranscribing: "Расшифровка аудио…",sectionTitle: "Диктовка",+ micTargetLabel: "Кнопка микрофона",+ micTargetAsr: "Распознавать через ASR",+ micTargetModel: "Отправлять аудио модели",+ micTargetAsrDescription: "Кнопка микрофона вводит сказанное в поле сообщения.",+ micTargetModelDescription: "Если загружена модель со звуковым вводом, микрофон записывает клип и прикрепляет его к сообщению. Для остальных моделей по-прежнему работает диктовка.",+ micTargetModelUnavailable: "Загрузите модель со звуковым вводом (например, LFM3-Omni), чтобы отправлять ей свой голос.",+ micRecordForModel: "Записать аудио для модели",engineLabel: "Движок диктовки",engineBrowser: "Браузер",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/sv.ts b/studio/frontend/src/i18n/locales/sv.tsindex 57574b9e1dc..cca9308ac90 100644--- a/studio/frontend/src/i18n/locales/sv.ts+++ b/studio/frontend/src/i18n/locales/sv.ts@@ -1766,6 +1766,13 @@ export const sv = {audioUploadCancel: "Avbryt transkribering",audioUploadTranscribing: "Transkriberar ljud…",sectionTitle: "Diktering",+ micTargetLabel: "Mikrofonknapp",+ micTargetAsr: "Transkribera med ASR",+ micTargetModel: "Skicka ljud till modellen",+ micTargetAsrDescription: "Mikrofonknappen skriver det du säger i meddelanderutan.",+ micTargetModelDescription: "När en modell med ljudinmatning är inläst spelar mikrofonen in ett klipp och bifogar det till ditt meddelande. Andra modeller får fortfarande diktering.",+ micTargetModelUnavailable: "Läs in en modell som tar emot ljudinmatning (till exempel LFM3-Omni) för att skicka din röst till den.",+ micRecordForModel: "Spela in ljud för modellen",engineLabel: "Dikteringsmotor",engineBrowser: "Webbläsare",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/th.ts b/studio/frontend/src/i18n/locales/th.tsindex 75ed6f51117..0f3a45eb1e8 100644--- a/studio/frontend/src/i18n/locales/th.ts+++ b/studio/frontend/src/i18n/locales/th.ts@@ -1748,6 +1748,13 @@ export const th = {audioUploadCancel: "ยกเลิกการถอดเสียง",audioUploadTranscribing: "กำลังถอดเสียง…",sectionTitle: "การป้อนตามคำบอก",+ micTargetLabel: "ปุ่มไมโครโฟน",+ micTargetAsr: "ถอดความด้วย ASR",+ micTargetModel: "ส่งเสียงไปยังโมเดล",+ micTargetAsrDescription: "ปุ่มไมโครโฟนจะพิมพ์สิ่งที่คุณพูดลงในช่องข้อความ",+ micTargetModelDescription: "เมื่อโหลดโมเดลที่รับอินพุตเสียง ไมโครโฟนจะบันทึกคลิปและแนบไปกับข้อความของคุณ โมเดลอื่นยังคงใช้การป้อนตามคำบอก",+ micTargetModelUnavailable: "โหลดโมเดลที่รับอินพุตเสียงได้ (เช่น LFM3-Omni) เพื่อส่งเสียงของคุณไปให้",+ micRecordForModel: "บันทึกเสียงสำหรับโมเดล",engineLabel: "เอนจินการป้อนตามคำบอก",engineBrowser: "เบราว์เซอร์",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/tr.ts b/studio/frontend/src/i18n/locales/tr.tsindex 3bad8ca43f4..e76577a510b 100644--- a/studio/frontend/src/i18n/locales/tr.ts+++ b/studio/frontend/src/i18n/locales/tr.ts@@ -1825,6 +1825,13 @@ export const tr = {audioUploadCancel: "Metne dönüştürmeyi iptal et",audioUploadTranscribing: "Ses metne dönüştürülüyor…",sectionTitle: "Dikte",+ micTargetLabel: "Mikrofon düğmesi",+ micTargetAsr: "ASR ile yazıya dök",+ micTargetModel: "Sesi modele gönder",+ micTargetAsrDescription: "Mikrofon düğmesi söylediklerinizi mesaj kutusuna yazar.",+ micTargetModelDescription: "Ses girişi alan bir model yüklüyken mikrofon bir klip kaydeder ve mesajınıza ekler. Diğer modeller dikte almaya devam eder.",+ micTargetModelUnavailable: "Sesinizi göndermek için ses girişi kabul eden bir model yükleyin (örneğin LFM3-Omni).",+ micRecordForModel: "Model için ses kaydet",engineLabel: "Dikte motoru",engineBrowser: "Tarayıcı",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/uk.ts b/studio/frontend/src/i18n/locales/uk.tsindex ae479bb6bc2..1e6b8e8cecf 100644--- a/studio/frontend/src/i18n/locales/uk.ts+++ b/studio/frontend/src/i18n/locales/uk.ts@@ -1840,6 +1840,13 @@ export const uk = {audioUploadCancel: "Скасувати транскрибування",audioUploadTranscribing: "Транскрибування аудіо…",sectionTitle: "Диктування",+ micTargetLabel: "Кнопка мікрофона",+ micTargetAsr: "Розпізнавати через ASR",+ micTargetModel: "Надсилати аудіо моделі",+ micTargetAsrDescription: "Кнопка мікрофона вводить сказане в поле повідомлення.",+ micTargetModelDescription: "Якщо завантажено модель зі звуковим введенням, мікрофон записує кліп і додає його до повідомлення. Для інших моделей і далі працює диктування.",+ micTargetModelUnavailable: "Завантажте модель зі звуковим введенням (наприклад, LFM3-Omni), щоб надсилати їй свій голос.",+ micRecordForModel: "Записати аудіо для моделі",engineLabel: "Рушій диктування",engineBrowser: "Браузер",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/vi.ts b/studio/frontend/src/i18n/locales/vi.tsindex 2d4177b32b9..933bdd96421 100644--- a/studio/frontend/src/i18n/locales/vi.ts+++ b/studio/frontend/src/i18n/locales/vi.ts@@ -1835,6 +1835,13 @@ export const vi = {audioUploadCancel: "Hủy chép lời",audioUploadTranscribing: "Đang chép lời âm thanh…",sectionTitle: "Đọc chính tả",+ micTargetLabel: "Nút micrô",+ micTargetAsr: "Chép lời bằng ASR",+ micTargetModel: "Gửi âm thanh cho mô hình",+ micTargetAsrDescription: "Nút micrô nhập những gì bạn nói vào ô tin nhắn.",+ micTargetModelDescription: "Khi đã tải mô hình nhận đầu vào âm thanh, micrô sẽ ghi một đoạn và đính kèm vào tin nhắn của bạn. Các mô hình khác vẫn dùng đọc chính tả.",+ micTargetModelUnavailable: "Hãy tải một mô hình nhận đầu vào âm thanh (ví dụ LFM3-Omni) để gửi giọng nói của bạn.",+ micRecordForModel: "Ghi âm cho mô hình",engineLabel: "Công cụ đọc chính tả",engineBrowser: "Trình duyệt",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.tsindex fbd02e46bf9..8bf3f5cc897 100644--- a/studio/frontend/src/i18n/locales/zh-CN.ts+++ b/studio/frontend/src/i18n/locales/zh-CN.ts@@ -1658,6 +1658,13 @@ export const zhCN = {audioUploadCancel: "取消转录",audioUploadTranscribing: "正在转录音频…",sectionTitle: "语音输入",+ micTargetLabel: "麦克风按钮",+ micTargetAsr: "使用 ASR 转写",+ micTargetModel: "将音频发送给模型",+ micTargetAsrDescription: "麦克风按钮会把你说的话输入到消息框中。",+ micTargetModelDescription: "加载支持音频输入的模型后,麦克风会录制一段音频并附加到你的消息中。其他模型仍使用语音输入。",+ micTargetModelUnavailable: "加载支持音频输入的模型(例如 LFM3-Omni)即可向它发送语音。",+ micRecordForModel: "为模型录制音频",engineLabel: "语音输入引擎",engineBrowser: "浏览器",engineBrowserDescription:diff --git a/studio/frontend/src/i18n/locales/zh-TW.ts b/studio/frontend/src/i18n/locales/zh-TW.tsindex b17e34dfe89..bacfbe42206 100644--- a/studio/frontend/src/i18n/locales/zh-TW.ts+++ b/studio/frontend/src/i18n/locales/zh-TW.ts@@ -1740,6 +1740,13 @@ export const zhTW = {audioUploadCancel: "取消轉錄",audioUploadTranscribing: "正在轉錄音訊…",sectionTitle: "聽寫",+ micTargetLabel: "麥克風按鈕",+ micTargetAsr: "使用 ASR 轉寫",+ micTargetModel: "將音訊傳送給模型",+ micTargetAsrDescription: "麥克風按鈕會把你說的話輸入到訊息框中。",+ micTargetModelDescription: "載入支援音訊輸入的模型後,麥克風會錄製一段音訊並附加到你的訊息中。其他模型仍使用聽寫。",+ micTargetModelUnavailable: "載入支援音訊輸入的模型(例如 LFM3-Omni)即可向它傳送語音。",+ micRecordForModel: "為模型錄製音訊",engineLabel: "聽寫引擎",engineBrowser: "瀏覽器",engineBrowserDescription:diff --git a/studio/frontend/tests/chat-audio-upload.test.ts b/studio/frontend/tests/chat-audio-upload.test.tsindex c15d364360a..0a8eb5716a1 100644--- a/studio/frontend/tests/chat-audio-upload.test.ts+++ b/studio/frontend/tests/chat-audio-upload.test.ts@@ -184,7 +184,8 @@ test("the main composer has one Dictate entry and does not gate it on draft text);const controlsStart = threadSource.indexOf("const ComposerRightControls:");const controls = threadSource.slice(controlsStart);- assert.equal((controls.match(/tooltip="Dictate"/g) ?? []).length, 1);+ assert.equal((controls.match(/tooltip=\{micLabel\}/g) ?? []).length, 1);+ assert.match(controls, /\? t\("settings\.voice\.dictation\.micRecordForModel"\)\s*: "Dictate";/);assert.doesNotMatch(controls, /Upload01Icon|DialogTrigger/);});diff --git a/studio/frontend/tests/mic-clip-recording.test.ts b/studio/frontend/tests/mic-clip-recording.test.tsnew file mode 100644index 00000000000..71cadf7f4c4--- /dev/null+++ b/studio/frontend/tests/mic-clip-recording.test.ts@@ -0,0 +1,410 @@+// 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 assert from "node:assert/strict";+import test, { mock } from "node:test";++import * as dictationOutcome from "../src/features/chat/adapters/dictation-outcome.ts";+import * as micClipState from "../src/features/chat/adapters/mic-clip-state.ts";+import * as micTarget from "../src/features/chat/utils/mic-target.ts";+import { loadWithStubs } from "./helpers/module-stubs.ts";++const MAX = 25 * 1024 * 1024;+const toasts: string[] = [];+let lastRecorder: FakeRecorder | null = null;++class FakeRecorder {+ state: RecordingState = "inactive";+ private data: ((event: { data: Blob }) => void)[] = [];+ constructor() {+ lastRecorder = this;+ }+ static rate = 16_000;+ sampleRate = FakeRecorder.rate;+ limitBytes = 0;+ bytesFor(seconds: number) {+ return 44 + Math.ceil(seconds * this.sampleRate) * 2;+ }+ maxBytes = Number.POSITIVE_INFINITY;+ secondsWithin(maxBytes: number) {+ this.limitBytes = maxBytes;+ return 30;+ }+ addEventListener(type: string, listener: (event: { data: Blob }) => void) {+ if (type === "dataavailable") this.data.push(listener);+ }+ start() {+ this.state = "recording";+ }+ stopCalls = 0;+ discarded = false;+ discard() {+ this.state = "inactive";+ this.discarded = true;+ }+ stop() {+ this.stopCalls += 1;+ this.state = "inactive";+ const data = new Blob([new Uint8Array(64)], { type: "audio/wav" });+ for (const listener of this.data) listener({ data });+ }+}++const stream = { getTracks: () => [{ stop() {} }] };+let micOpens = 0;+let micDeviceId = "default";+Object.defineProperty(globalThis, "navigator", {+ value: {+ mediaDevices: {+ getUserMedia: async () => {+ micOpens += 1;+ return stream;+ },+ },+ },+ configurable: true,+});++const { discardClipOnRelease, discardClipWhenModelChanges, listenForAudioClip } =+ loadWithStubs<+ typeof import("../src/features/chat/adapters/studio-audio-clip-adapter.ts")+>(+ new URL("../src/features/chat/adapters/studio-audio-clip-adapter.ts", import.meta.url),+ {+ "@/features/settings/stores/voice-settings-store": {+ useVoiceSettingsStore: { getState: () => ({ micDeviceId }) },+ },+ "@/lib/audio-utils": { MAX_AUDIO_SIZE: MAX, MAX_AUDIO_SIZE_LABEL: "25MB" },+ sonner: {+ toast: {+ info: (message: string) => toasts.push(message),+ error: (message: string) => toasts.push(message),+ },+ },+ "../utils/mic-target": micTarget,+ "./dictation-level": { startDictationLevelMeter: () => () => {} },+ "./dictation-outcome": dictationOutcome,+ "./mic-clip-state": micClipState,+ "./pcm-recorder": { PcmRecorder: FakeRecorder },+ "./studio-web-speech-dictation-adapter": {+ isMissingDeviceError: (error: unknown) =>+ (error as Error)?.name === "NotFoundError",+ },+ },+);++const flush = () => new Promise((resolve) => setImmediate(resolve));++test("a recording stops itself at the audio attachment limit and keeps the clip", async () => {+ mock.timers.enable({ apis: ["setTimeout"] });+ try {+ const clips: File[] = [];+ const session = listenForAudioClip((file) => clips.push(file));+ await flush();+ assert.equal(session.status.type, "running");+ assert.equal(lastRecorder?.limitBytes, MAX);+ mock.timers.tick(29_999);+ assert.equal(lastRecorder?.state, "recording");+ mock.timers.tick(1);+ await flush();+ assert.equal(lastRecorder?.state, "inactive");+ assert.equal(clips.length, 1);+ assert.equal(clips[0].type, "audio/wav");+ assert.equal(session.status.type, "ended");+ assert.equal(dictationOutcome.dictationProducedClip(), clips[0]);+ assert.deepEqual(toasts, ["Recording stopped at the 25MB audio limit."]);+ } finally {+ mock.timers.reset();+ }+});++test("a stop before the limit clears the auto-stop", async () => {+ mock.timers.enable({ apis: ["setTimeout"] });+ try {+ toasts.length = 0;+ const clips: File[] = [];+ const session = listenForAudioClip((file) => clips.push(file));+ await flush();+ await session.stop();+ mock.timers.tick(60_000);+ await flush();+ assert.equal(clips.length, 1);+ assert.deepEqual(toasts, []);+ } finally {+ mock.timers.reset();+ }+});++test("cancel discards the recording", async () => {+ const clips: File[] = [];+ const session = listenForAudioClip((file) => clips.push(file));+ await flush();+ session.cancel();+ await flush();+ assert.equal(clips.length, 0);+ assert.equal(lastRecorder?.stopCalls, 0);+ assert.equal(lastRecorder?.discarded, true);+ assert.deepEqual(session.status, { type: "ended", reason: "cancelled" });+});++test("a recording stops within the room left by earlier attachments", async () => {+ const session = listenForAudioClip(() => {}, { maxBytes: 1_000_000 });+ await flush();+ assert.equal(lastRecorder?.limitBytes, 1_000_000);+ session.cancel();+});++test("no room left refuses before opening the mic", async () => {+ toasts.length = 0;+ lastRecorder = null;+ const opensBefore = micOpens;+ let ended = false;+ const session = listenForAudioClip(() => {}, { reason: "No room." });+ session.onEnd?.(() => {+ ended = true;+ });+ await flush();+ assert.equal(micOpens, opensBefore);+ assert.equal(lastRecorder, null);+ assert.equal(ended, true);+ assert.deepEqual(session.status, { type: "ended", reason: "error" });+ assert.deepEqual(toasts, ["No room."]);+});++test("other audio is held off while recording, and the clip is capped at the room", async () => {+ let recordingAtClip: boolean | null = null;+ const session = listenForAudioClip(+ () => {+ recordingAtClip = micClipState.micClipRecording();+ },+ { maxBytes: 500_000 },+ );+ assert.equal(micClipState.micClipRecording(), true);+ await flush();+ assert.equal(lastRecorder?.maxBytes, 500_000);+ await session.stop();+ assert.equal(recordingAtClip, false);+ assert.equal(micClipState.micClipRecording(), false);+});++test("cancel and a refused start leave other audio free to attach", async () => {+ const session = listenForAudioClip(() => {});+ await flush();+ session.cancel();+ assert.equal(micClipState.micClipRecording(), false);+ listenForAudioClip(() => {}, { reason: "No room." });+ assert.equal(micClipState.micClipRecording(), false);+ await flush();+});++test("a stale session's late cleanup keeps a newer recording's hold", async () => {+ let resolveFirst: (value: unknown) => void = () => {};+ const nav = globalThis.navigator as unknown as {+ mediaDevices: { getUserMedia: () => Promise<unknown> };+ };+ const original = nav.mediaDevices.getUserMedia;+ nav.mediaDevices.getUserMedia = () =>+ new Promise((resolve) => {+ resolveFirst = resolve;+ });+ const first = listenForAudioClip(() => {});+ first.cancel();+ nav.mediaDevices.getUserMedia = original;+ const second = listenForAudioClip(() => {});+ await flush();+ assert.equal(micClipState.micClipRecording(), true);+ // The first mic request resolves after the user already started again.+ resolveFirst(stream);+ await flush();+ assert.equal(micClipState.micClipRecording(), true);+ second.cancel();+ assert.equal(micClipState.micClipRecording(), false);+});++test("a cancel during the saved-mic request skips the default-mic fallback", async () => {+ const nav = globalThis.navigator as unknown as {+ mediaDevices: { getUserMedia: () => Promise<unknown> };+ };+ const original = nav.mediaDevices.getUserMedia;+ let rejectSaved: (error: Error) => void = () => {};+ let requests = 0;+ nav.mediaDevices.getUserMedia = () => {+ requests += 1;+ return requests === 1+ ? new Promise((_resolve, reject) => {+ rejectSaved = reject;+ })+ : Promise.resolve(stream);+ };+ micDeviceId = "unplugged-usb-mic";+ toasts.length = 0;+ try {+ const session = listenForAudioClip(() => {});+ session.cancel();+ rejectSaved(Object.assign(new Error("gone"), { name: "NotFoundError" }));+ await flush();+ assert.equal(requests, 1);+ assert.deepEqual(toasts, []);+ assert.deepEqual(session.status, { type: "ended", reason: "cancelled" });+ } finally {+ nav.mediaDevices.getUserMedia = original;+ micDeviceId = "default";+ }+});++test("a cancelled session's late mic error leaves the next session's outcome alone", async () => {+ const nav = globalThis.navigator as unknown as {+ mediaDevices: { getUserMedia: () => Promise<unknown> };+ };+ const original = nav.mediaDevices.getUserMedia;+ let rejectFirst: (error: Error) => void = () => {};+ let requests = 0;+ nav.mediaDevices.getUserMedia = () => {+ requests += 1;+ return requests === 1+ ? new Promise((_resolve, reject) => {+ rejectFirst = reject;+ })+ : Promise.resolve(stream);+ };+ toasts.length = 0;+ try {+ const first = listenForAudioClip(() => {});+ first.cancel();+ const second = listenForAudioClip(() => {});+ await flush();+ assert.equal(second.status.type, "running");+ rejectFirst(Object.assign(new Error("busy"), { name: "NotReadableError" }));+ await flush();+ assert.equal(dictationOutcome.dictationFailed(), false);+ assert.deepEqual(toasts, []);+ second.cancel();+ } finally {+ nav.mediaDevices.getUserMedia = original;+ }+});++function fakeCheckpointStore(checkpoint: string) {+ let state: {+ params: { checkpoint: string };+ modelLoading?: boolean;+ loadingModelPick?: unknown;+ } = { params: { checkpoint } };+ const listeners = new Set<(s: typeof state) => void>();+ return {+ getState: () => state,+ subscribe: (listener: (s: typeof state) => void) => {+ listeners.add(listener);+ return () => listeners.delete(listener);+ },+ set(next: string, extra: { modelLoading?: boolean; loadingModelPick?: unknown } = {}) {+ state = { params: { checkpoint: next }, ...extra };+ for (const listener of [...listeners]) listener(state);+ },+ listeners,+ };+}++test("switching the chat model mid-recording discards the clip with a toast", async () => {+ toasts.length = 0;+ const clips: File[] = [];+ const store = fakeCheckpointStore("gemma-4-e2b");+ const session = listenForAudioClip((file) => clips.push(file));+ discardClipWhenModelChanges(session, store);+ await flush();+ store.set("gemma-4-e2b");+ assert.equal(session.status.type, "running");+ store.set("qwen3-text-only");+ await flush();+ assert.deepEqual(session.status, { type: "ended", reason: "cancelled" });+ assert.equal(lastRecorder?.discarded, true);+ assert.equal(clips.length, 0);+ assert.deepEqual(toasts, ["Recording discarded because the chat model changed."]);+ assert.equal(store.listeners.size, 0);+});++test("a finished recording stops watching the chat model", async () => {+ toasts.length = 0;+ const store = fakeCheckpointStore("gemma-4-e2b");+ const session = listenForAudioClip(() => {});+ discardClipWhenModelChanges(session, store);+ await flush();+ await session.stop();+ assert.equal(store.listeners.size, 0);+ store.set("qwen3-text-only");+ assert.deepEqual(toasts, []);+});++test("a model load starting mid-recording discards the clip before checkpoint moves", async () => {+ for (const extra of [+ { modelLoading: true },+ { loadingModelPick: { id: "qwen3-text-only" } },+ ]) {+ toasts.length = 0;+ const store = fakeCheckpointStore("gemma-4-e2b");+ const session = listenForAudioClip(() => {});+ discardClipWhenModelChanges(session, store);+ await flush();+ store.set("gemma-4-e2b", extra);+ await flush();+ assert.deepEqual(session.status, { type: "ended", reason: "cancelled" });+ assert.deepEqual(toasts, ["Recording discarded because the chat model changed."]);+ }+});++test("a room sized for 16 kHz is refused when the recorder opens at device rate", async () => {+ toasts.length = 0;+ FakeRecorder.rate = 48_000;+ try {+ const clips: File[] = [];+ const session = listenForAudioClip((file) => clips.push(file), {+ maxBytes: 44 + 16_000 * 2,+ });+ await flush();+ assert.deepEqual(session.status, { type: "ended", reason: "error" });+ assert.equal(lastRecorder?.discarded, true);+ assert.equal(lastRecorder?.stopCalls, 0);+ assert.equal(clips.length, 0);+ assert.deepEqual(toasts, ["Audio files together exceed the 25MB per-message limit"]);+ assert.equal(micClipState.micClipRecording(), false);+ } finally {+ FakeRecorder.rate = 16_000;+ }+});++test("the same room is enough at 16 kHz", async () => {+ const session = listenForAudioClip(() => {}, { maxBytes: 44 + 16_000 * 2 });+ await flush();+ assert.equal(session.status.type, "running");+ session.cancel();+});++test("switching chats mid-recording discards the clip instead of attaching it elsewhere", async () => {+ toasts.length = 0;+ const clips: File[] = [];+ const onRelease = new Set<() => void>();+ const session = listenForAudioClip((file) => clips.push(file));+ discardClipOnRelease(session, onRelease);+ await flush();+ assert.equal(session.status.type, "running");+ for (const listener of [...onRelease]) listener();+ await flush();+ assert.deepEqual(session.status, { type: "ended", reason: "cancelled" });+ assert.equal(lastRecorder?.discarded, true);+ assert.equal(clips.length, 0);+ assert.deepEqual(toasts, ["Recording discarded because the chat changed."]);+ assert.equal(onRelease.size, 0);+});++test("a finished recording is not discarded by a later chat switch", async () => {+ toasts.length = 0;+ const clips: File[] = [];+ const onRelease = new Set<() => void>();+ const session = listenForAudioClip((file) => clips.push(file));+ discardClipOnRelease(session, onRelease);+ await flush();+ await session.stop();+ assert.equal(clips.length, 1);+ assert.equal(onRelease.size, 0);+ assert.deepEqual(toasts, []);+});diff --git a/studio/frontend/tests/mic-target.test.ts b/studio/frontend/tests/mic-target.test.tsnew file mode 100644index 00000000000..16cc8c08697--- /dev/null+++ b/studio/frontend/tests/mic-target.test.ts@@ -0,0 +1,277 @@+// 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 assert from "node:assert/strict";+import test from "node:test";+import * as zustand from "zustand";+import * as zustandMiddleware from "zustand/middleware";++import * as audioCppCatalog from "../src/features/audio/audio-cpp-catalog.ts";+import {+ beginDictationSession,+ dictationProducedClip,+ markDictationClip,+} from "../src/features/chat/adapters/dictation-outcome.ts";+import {+ clipStillAttached,+ shouldSubmitDictation,+} from "../src/features/chat/utils/dictation-send.ts";+import {+ micClipFilename,+ MIN_MIC_CLIP_BYTES,+ micClipRoom,+ micSendsAudioToModel,+ pickAudioForModel,+} from "../src/features/chat/utils/mic-target.ts";+import { AUDIO_ATTACHMENT_ACCEPT, MAX_AUDIO_SIZE } from "../src/lib/audio-utils.ts";+import * as sttModelCatalog from "../src/features/settings/stores/stt-model-catalog.ts";+import { installLocalStorageFake, readSrc } from "./helpers/kit.ts";+import { loadWithStubs } from "./helpers/module-stubs.ts";++const { store } = installLocalStorageFake();+const KEY = "unsloth_voice_settings";++const { useVoiceSettingsStore } = loadWithStubs<+ typeof import("../src/features/settings/stores/voice-settings-store.ts")+>(+ new URL("../src/features/settings/stores/voice-settings-store.ts", import.meta.url),+ {+ "@/lib/api-base": { isTauri: false },+ "../../audio/audio-cpp-catalog": audioCppCatalog,+ zustand,+ "zustand/middleware": zustandMiddleware,+ "./stt-model-catalog": sttModelCatalog,+ "../api/stt-device": { syncSttDevicePreference: () => undefined },+ },+);++const rehydrateWith = async (saved: Record<string, unknown>) => {+ store.set(KEY, JSON.stringify({ state: saved, version: 1 }));+ await useVoiceSettingsStore.persist.rehydrate();+};++test("the mic defaults to ASR dictation", () => {+ assert.equal(useVoiceSettingsStore.getInitialState().micTarget, "asr");+});++test("an older saved settings blob without the field stays on ASR", async () => {+ await rehydrateWith({ dictationEngine: "model" });+ assert.equal(useVoiceSettingsStore.getState().micTarget, "asr");+});++test("a saved model pick survives a reload, junk falls back to ASR", async () => {+ await rehydrateWith({ micTarget: "model" });+ assert.equal(useVoiceSettingsStore.getState().micTarget, "model");+ await rehydrateWith({ micTarget: "nonsense" });+ assert.equal(useVoiceSettingsStore.getState().micTarget, "asr");+});++test("ASR never sends audio to the model, even one that hears it", () => {+ assert.equal(micSendsAudioToModel("asr", true), false);+});++test("model mode needs a loaded model with audio input", () => {+ assert.equal(micSendsAudioToModel("model", true), true);+ assert.equal(micSendsAudioToModel("model", false), false);+});++test("the settings option is disabled without an audio-input model", () => {+ const source = readSrc("features/settings/tabs/voice-tab.tsx");+ assert.match(+ source,+ /value="model"\s+data-testid="mic-target-model"\s+disabled=\{!modelHearsAudio && micTarget !== "model"\}/,+ );+ assert.match(source, /\?\.hasAudioInput === true/);+});++test("Compare and the adapter route only through the gate", () => {+ const source = readSrc("features/chat/adapters/studio-dictation-adapter.tsx");+ assert.match(source, /if \(chatId === null \|\| !audioClipTarget\) return false;/);+ assert.match(+ source,+ /micSendsAudioToModel\(micTarget, activeModel\?\.hasAudioInput === true\)/,+ );+ assert.match(+ source,+ /if \(clipTarget && micRecordsForModel\(this\.chatId\)\) \{[\s\S]*?listenForAudioClip\(clipTarget\.add, room\);\s*discardClipWhenModelChanges\(session, useChatRuntimeStore\)/,+ ); assert.match(source, /discardClipOnRelease\(session, clipTarget\.onRelease\)/);+ const thread = readSrc("components/assistant-ui/thread.tsx");+ const register = thread.slice(thread.indexOf("setMicAudioClipTarget({"));+ assert.match(register.slice(0, 700), /\[aui, composerIdentity\]/);+});++test("a recorded clip is a WAV named for the attachment chip", () => {+ assert.equal(+ micClipFilename(new Date("2026-11-04T18:05:09.123Z")),+ "recording-2026-11-04T18-05-09.wav",+ );+});++test("a clip marks the session and resets on the next one", () => {+ beginDictationSession();+ assert.equal(dictationProducedClip(), null);+ const clip = new File([new Uint8Array(4)], "recording.wav", { type: "audio/wav" });+ markDictationClip(clip);+ assert.equal(dictationProducedClip(), clip);+ beginDictationSession();+ assert.equal(dictationProducedClip(), null);+});++test("a held stop-and-send drops the clip once it is removed from the composer", () => {+ const clip = new File([new Uint8Array(4)], "recording.wav", { type: "audio/wav" });+ const other = new File([new Uint8Array(4)], "recording.wav", { type: "audio/wav" });+ assert.equal(clipStillAttached(clip, [{ file: clip }]), true);+ assert.equal(clipStillAttached(clip, [{ file: other }, {}]), false);+ assert.equal(clipStillAttached(clip, []), false);+ assert.equal(clipStillAttached(null, [{ file: clip }]), false);+ const base = { originComposer: "a", currentComposer: "a", producedTranscript: false, baseText: "", text: "hi" };+ assert.equal(shouldSubmitDictation({ ...base, producedClip: clipStillAttached(clip, []) }), false);+ const thread = readSrc("components/assistant-ui/thread.tsx");+ assert.match(+ thread,+ /producedClip: clipStillAttached\(\s*dictationProducedClip\(\),\s*aui\.composer\(\)\.getState\(\)\.attachments,\s*\)/,+ );+});++test("stop-and-send submits an attached clip with no new text", () => {+ const base = {+ originComposer: "a",+ currentComposer: "a",+ producedTranscript: false,+ baseText: "",+ text: "",+ };+ assert.equal(shouldSubmitDictation({ ...base, producedClip: true }), true);+ assert.equal(shouldSubmitDictation(base), false);+ assert.equal(+ shouldSubmitDictation({ ...base, producedClip: true, currentComposer: "b" }),+ false,+ );+});++// Every route into the store must carry hasAudioInput: list row, load response, status adoption.+test("the gate is family-agnostic: no model names in the routing", () => {+ const gate = readSrc("features/chat/adapters/studio-dictation-adapter.tsx");+ const routing = gate.slice(gate.indexOf("export function micRecordsForModel"));+ assert.doesNotMatch(routing.slice(0, routing.indexOf("\n}\n")), /lfm|gemma|qwen|omni|audio_type/i);+});++test("every route into the chat store carries has_audio_input", () => {+ const runtime = readSrc("features/chat/hooks/use-chat-model-runtime.ts");+ const row = runtime.slice(runtime.indexOf("function toChatModelRow"));+ assert.match(row.slice(0, 900), /hasAudioInput: Boolean\(model\.has_audio_input\)/);+ const sync = runtime.slice(runtime.indexOf("export function syncModelCapabilities"));+ assert.match(sync.slice(0, 2000), /hasAudioInput: Boolean\(resp\.has_audio_input\)/);+ const status = readSrc("features/chat/lib/apply-inference-status-to-store.ts");+ assert.match(status, /hasAudioInput: status\.has_audio_input \?\? false/);+});++test("a recording gets the whole audio limit when nothing else is attached", () => {+ assert.deepEqual(micClipRoom([], 8, false), { maxBytes: MAX_AUDIO_SIZE });+ const image = { name: "a.png", type: "image/png", size: 5_000_000 };+ assert.deepEqual(micClipRoom([image], 8, false), { maxBytes: MAX_AUDIO_SIZE });+});++test("a recording is capped by the audio already in the composer", () => {+ const clip = { name: "a.wav", type: "audio/wav", size: 10_000_000 };+ const m4a = { name: "b.m4a", type: "", size: 5_000_000 };+ assert.deepEqual(micClipRoom([clip, m4a], 8, false), {+ maxBytes: MAX_AUDIO_SIZE - 15_000_000,+ });+});++test("no room is refused with the attachment adapter's reason", () => {+ const clip = { name: "a.wav", type: "audio/wav", size: 1_000 };+ assert.match(+ (micClipRoom([clip], 1, false) as { reason: string }).reason,+ /one audio file per message/,+ );+ const full = { name: "a.wav", type: "audio/wav", size: MAX_AUDIO_SIZE };+ assert.match(+ (micClipRoom([full], 8, false) as { reason: string }).reason,+ /per-message limit/,+ );+ assert.match(+ (micClipRoom([], 8, true) as { reason: string }).reason,+ /staged audio clip/,+ );+});++test("a room too small for a usable clip is refused", () => {+ const nearlyFull = {+ name: "a.wav",+ type: "audio/wav",+ size: MAX_AUDIO_SIZE - MIN_MIC_CLIP_BYTES + 1,+ };+ assert.match(+ (micClipRoom([nearlyFull], 8, false) as { reason: string }).reason,+ /per-message limit/,+ );+ const justFits = { ...nearlyFull, size: MAX_AUDIO_SIZE - MIN_MIC_CLIP_BYTES };+ assert.deepEqual(micClipRoom([justFits], 8, false), {+ maxBytes: MIN_MIC_CLIP_BYTES,+ });+});++test("the audio attachment adapter refuses audio while a mic clip records", () => {+ const adapter = readSrc("features/chat/audio-attachment-adapter.ts");+ assert.match(+ adapter,+ /if \(micClipRecording\(\)\) \{[\s\S]*?throw new Error\(recordingReason\)/,+ );+});++test("the model clip path gates on PCM support, not MediaRecorder", () => {+ const source = readSrc("features/chat/adapters/studio-dictation-adapter.tsx");+ assert.match(+ source,+ /if \(clipTarget && micRecordsForModel\(this\.chatId\)\) \{\s*if \(!PcmRecorder\.isSupported\(\)\)/,+ );+});++test("over plain HTTP the model mic picks recording files and attaches them", async () => {+ const listeners = new Map<string, () => void>();+ const input = {+ type: "",+ accept: "",+ multiple: false,+ hidden: false,+ files: null as File[] | null,+ removed: false,+ clicked: false,+ addEventListener: (type: string, listener: () => void) => listeners.set(type, listener),+ remove() {+ this.removed = true;+ },+ click() {+ this.clicked = true;+ },+ };+ let appended: unknown = null;+ const doc = {+ createElement: () => input,+ body: { append: (node: unknown) => (appended = node) },+ } as unknown as Document;+ const added: File[] = [];+ const clip = new File([new Uint8Array(4)], "note.m4a", { type: "" });+ const refused = new File([new Uint8Array(4)], "big.wav", { type: "audio/wav" });+ pickAudioForModel(async (file) => {+ added.push(file);+ if (file === refused) throw new Error("too big");+ }, doc);+ assert.equal(input.type, "file");+ assert.equal(input.accept, AUDIO_ATTACHMENT_ACCEPT);+ assert.equal(appended, input);+ assert.equal(input.clicked, true);+ input.files = [clip, refused];+ listeners.get("change")?.();+ await new Promise((resolve) => setImmediate(resolve));+ assert.deepEqual(added, [clip, refused]);+ assert.equal(input.removed, true);+ const thread = readSrc("components/assistant-ui/thread.tsx");+ const fallback = thread.slice(thread.indexOf('currentDictationEntryMode() === "recording-file"'));+ assert.match(+ fallback.slice(0, 400),+ /if \(micRecordsForModel\(\)\) \{\s*pickAudioForModel\(\(file\) => aui\.composer\(\)\.addAttachment\(file\)\);\s*\} else \{\s*audioUpload\.openDialog\(\);/,+ );+});diff --git a/studio/frontend/tests/pcm-recorder.test.ts b/studio/frontend/tests/pcm-recorder.test.tsindex be1686e3c93..18a6a23b143 100644--- a/studio/frontend/tests/pcm-recorder.test.ts+++ b/studio/frontend/tests/pcm-recorder.test.ts@@ -403,3 +403,93 @@ test("the recording call sites go through createAudioRecorder", () => {assert.equal(source.includes("createAudioRecorder("), true, path);}});++test("discard tears the graph down without encoding or firing events", () => {+ const recorder = newRecorder();+ const events: string[] = [];+ recorder.addEventListener("dataavailable", () => events.push("dataavailable"));+ recorder.addEventListener("stop", () => events.push("stop"));+ recorder.start();+ FakeAudioContext.last?.processor.emit(new Float32Array([0.5, 0.5]));+ recorder.discard();+ assert.deepEqual(events, []);+ assert.equal(recorder.state, "inactive");+ assert.equal(FakeAudioContext.last?.closed, true);+ recorder.stop();+ assert.deepEqual(events, []);+});++test("stop trims the WAV to maxBytes", async () => {+ const recorder = newRecorder();+ let recorded: { data: Blob } | null = null;+ recorder.addEventListener("dataavailable", (event) => {+ recorded = event;+ });+ recorder.maxBytes = 44 + 3 * 2;+ recorder.start();+ FakeAudioContext.last?.processor.emit(new Float32Array([0.5, 0.5]));+ FakeAudioContext.last?.processor.emit(new Float32Array([0.5, 0.5]));+ recorder.stop();+ assert.ok(recorded);+ assert.equal((recorded as { data: Blob }).data.size, 44 + 3 * 2);+});++test("PCM recording is supported without MediaRecorder", () => {+ const saved = {+ window: Object.getOwnPropertyDescriptor(globalThis, "window"),+ navigator: Object.getOwnPropertyDescriptor(globalThis, "navigator"),+ MediaRecorder: Object.getOwnPropertyDescriptor(globalThis, "MediaRecorder"),+ };+ const define = (key: string, value: unknown) =>+ Object.defineProperty(globalThis, key, { value, configurable: true, writable: true });+ try {+ define("MediaRecorder", undefined);+ define("navigator", { mediaDevices: { getUserMedia: async () => ({}) } });+ define("window", { isSecureContext: true, AudioContext: FakeAudioContext });+ assert.equal(PcmRecorder.isSupported(), true);+ define("window", {+ isSecureContext: true,+ webkitAudioContext: FakeAudioContext,+ });+ assert.equal(PcmRecorder.isSupported(), true);+ define("window", { isSecureContext: false, AudioContext: FakeAudioContext });+ assert.equal(PcmRecorder.isSupported(), false);+ define("window", { isSecureContext: true });+ assert.equal(PcmRecorder.isSupported(), false);+ } finally {+ for (const [key, descriptor] of Object.entries(saved)) {+ if (descriptor) Object.defineProperty(globalThis, key, descriptor);+ else delete (globalThis as Record<string, unknown>)[key];+ }+ }+});++test("bytesFor sizes a WAV at the recorder's actual rate", () => {+ FakeAudioContext.refuseRequestedRate = true;+ try {+ const recorder = newRecorder();+ assert.equal(recorder.sampleRate, 48_000);+ assert.equal(recorder.bytesFor(1), 44 + 96_000);+ } finally {+ FakeAudioContext.refuseRequestedRate = false;+ }+});++test("discard closes a recorder that was never started, once", () => {+ const recorder = newRecorder();+ const context = FakeAudioContext.last;+ let closes = 0;+ if (context) {+ const close = context.close.bind(context);+ context.close = () => {+ closes += 1;+ return close();+ };+ }+ recorder.discard();+ assert.equal(context?.closed, true);+ assert.equal(context?.processor.disconnected, true);+ recorder.discard();+ recorder.stop();+ assert.equal(closes, 1);+});