feat: integrate Meta Muse Voice Transcribe realtime and async

Frontend/client/src/i18n/translations/de/settings.ts

@@ -1,6 +1,12 @@
 import type { TranslationCatalog } from "@/i18n/types";
 
 export const settingsTranslations = {
+  "Live partials and completed turns · up to 60 minutes · Meta Model API key":
+    "Live-Zwischenstände und abgeschlossene Äußerungen · bis 60 Minuten · Meta-Model-API-Schlüssel",
+  "Transcribes after stop · files up to 10 minutes · Meta Model API key":
+    "Transkription nach dem Stoppen · Dateien bis 10 Minuten · Meta-Model-API-Schlüssel",
+  "Native speaker labels and approximate turn timestamps. Maximum 10 minutes per recording.":
+    "Native Sprechertrennung und ungefähre Zeitstempel je Äußerung. Maximal 10 Minuten pro Aufnahme.",
   // Languages and shared settings states.
   "Auto-detect": "Automatisch erkennen",
   German: "Deutsch",

Frontend/client/src/lib/meta-transcription-settings.test.ts

@@ -0,0 +1,44 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+
+import { settingsTranslations } from "@/i18n/translations/de/settings";
+import {
+  META_CREDENTIAL_REQUIREMENT,
+  META_MEETING_FINAL_STT_OPTION,
+  META_TRANSCRIPTION_OPTIONS,
+  metaFrontendModelForService,
+  metaSettingsPatchForModel,
+} from "@/lib/meta-transcription-settings";
+
+test("Meta offers separate realtime and async routes for one published model", () => {
+  assert.deepEqual(
+    META_TRANSCRIPTION_OPTIONS.map((option) => option.group),
+    ["cloud_streaming", "cloud_async"],
+  );
+  for (const option of META_TRANSCRIPTION_OPTIONS) {
+    assert.equal(option.model, "muse-voice-transcribe-1.0");
+    assert.equal(option.usdPerHour, 0.18);
+    assert.equal(metaFrontendModelForService(option.service), option.value);
+    assert.deepEqual(metaSettingsPatchForModel(option.value), { defaultSttService: option.service });
+    assert.ok(settingsTranslations[option.routeNote]);
+    assert.equal("wordErrorRatePercent" in option, false);
+  }
+  assert.equal(metaSettingsPatchForModel("meta-unknown"), null);
+  assert.equal(metaFrontendModelForService("meta"), null);
+});
+
+test("both Meta STT routes reuse the existing Meta Model API credential", () => {
+  assert.deepEqual(META_CREDENTIAL_REQUIREMENT, {
+    provider: "Meta Model API",
+    label: "Meta Model API key",
+    helpKey: "meta",
+  });
+});
+
+test("Meta Meeting option does not promise five-hour or word-level support", () => {
+  assert.equal(META_MEETING_FINAL_STT_OPTION.value, "meta_stt_async");
+  assert.equal(META_MEETING_FINAL_STT_OPTION.nativeDiarization, true);
+  assert.equal(META_MEETING_FINAL_STT_OPTION.fiveHourSupported, false);
+  assert.match(META_MEETING_FINAL_STT_OPTION.detail, /10 minutes/);
+  assert.ok(settingsTranslations[META_MEETING_FINAL_STT_OPTION.detail]);
+});

Frontend/client/src/lib/meta-transcription-settings.ts

@@ -0,0 +1,49 @@
+// Both transports use the same published model; Async is not a second model.
+export const META_TRANSCRIPTION_OPTIONS = [
+  {
+    value: "meta-realtime",
+    service: "meta_stt",
+    label: "Meta Muse Voice Transcribe Realtime",
+    model: "muse-voice-transcribe-1.0",
+    group: "cloud_streaming",
+    icon: "meta",
+    usdPerHour: 0.18,
+    routeNote: "Live partials and completed turns · up to 60 minutes · Meta Model API key",
+  },
+  {
+    value: "meta-async",
+    service: "meta_stt_async",
+    label: "Meta Muse Voice Transcribe Async",
+    model: "muse-voice-transcribe-1.0",
+    group: "cloud_async",
+    icon: "meta",
+    usdPerHour: 0.18,
+    routeNote: "Transcribes after stop · files up to 10 minutes · Meta Model API key",
+  },
+] as const;
+
+export const META_CREDENTIAL_REQUIREMENT = {
+  provider: "Meta Model API",
+  label: "Meta Model API key",
+  helpKey: "meta",
+} as const;
+
+export const META_MEETING_FINAL_STT_OPTION = {
+  value: "meta_stt_async",
+  label: "Meta Muse Voice Transcribe",
+  model: "muse-voice-transcribe-1.0",
+  credentialModel: "meta-async",
+  recommended: false,
+  nativeDiarization: true,
+  fiveHourSupported: false,
+  detail: "Native speaker labels and approximate turn timestamps. Maximum 10 minutes per recording.",
+} as const;
+
+export function metaFrontendModelForService(service: string) {
+  return META_TRANSCRIPTION_OPTIONS.find((option) => option.service === service)?.value ?? null;
+}
+
+export function metaSettingsPatchForModel(value: string) {
+  const option = META_TRANSCRIPTION_OPTIONS.find((candidate) => candidate.value === value);
+  return option ? { defaultSttService: option.service } : null;
+}

Frontend/client/src/pages/Settings.tsx

@@ -150,6 +150,13 @@ import {
   geminiFrontendModelForService,
   geminiSettingsPatchForModel,
 } from "@/lib/gemini-transcription-settings";
+import {
+  META_TRANSCRIPTION_OPTIONS,
+  META_CREDENTIAL_REQUIREMENT,
+  META_MEETING_FINAL_STT_OPTION,
+  metaFrontendModelForService,
+  metaSettingsPatchForModel,
+} from "@/lib/meta-transcription-settings";
 
 type Translate = ReturnType<typeof useI18n>["t"];
 type FormatDate = ReturnType<typeof useI18n>["formatDate"];
@@ -255,6 +262,7 @@ const TRANSCRIPTION_MODEL_OPTIONS = [
   { value: "soniox-async", label: "Soniox Async" },
   { value: "modulate-realtime", label: "Modulate.AI Multilingual Realtime" },
   { value: "modulate-async", label: "Modulate.AI Multilingual Batch" },
+  ...META_TRANSCRIPTION_OPTIONS.map(({ value, label }) => ({ value, label })),
   {
     value: GEMINI_REALTIME_TRANSCRIPTION_OPTION.value,
     label: GEMINI_REALTIME_TRANSCRIPTION_OPTION.label,
@@ -892,6 +900,13 @@ function createProviderModelOptions(
 
   return [
     benchmarkOption("elevenlabs", "ElevenLabs Live", 6.5, 3.6, "cloud_streaming", "elevenlabs"),
+    ...META_TRANSCRIPTION_OPTIONS.map((option) => ({
+      ...option,
+      model: providerModels[option.value] || option.model,
+      detail: `${formatEstimatedEuroFromUsd(option.usdPerHour, localeTag)}/h`,
+      routeNote: t(option.routeNote),
+      hourlyCostEur: option.usdPerHour * USD_TO_EUR_FOR_ESTIMATES,
+    })),
     benchmarkOption(
       GEMINI_REALTIME_TRANSCRIPTION_OPTION.value,
       GEMINI_REALTIME_TRANSCRIPTION_OPTION.label,
@@ -974,6 +989,7 @@ function createProviderModelOptions(
 }
 
 const MEETING_FINAL_STT_OPTIONS = [
+  META_MEETING_FINAL_STT_OPTION,
   {
     value: "soniox_async",
     label: "Soniox Async",
@@ -2556,6 +2572,9 @@ export default function Settings() {
 
   const requiredCredentialForTranscriptionModel = (model: string): CredentialRequirement | null => {
     switch (model) {
+      case "meta-realtime":
+      case "meta-async":
+        return META_CREDENTIAL_REQUIREMENT;
       case "soniox-realtime":
       case "soniox-async":
         return { provider: "Soniox", label: "Soniox API key", helpKey: "soniox" };
@@ -2760,6 +2779,10 @@ export default function Settings() {
       if (service === "modulate" || service === "modulate_async") {
         return service === "modulate_async" ? "modulate-async" : "modulate-realtime";
       }
+      const metaModel = metaFrontendModelForService(service);
+      if (metaModel) {
+        return metaModel;
+      }
       if (service === "mistral" || service === "mistral_async") {
         return service === "mistral_async" ? "mistral-async" : "mistral-realtime";
       }
@@ -3320,6 +3343,7 @@ export default function Settings() {
     }
     const previousValue = transcriptionModel;
     const geminiSettingsPatch = geminiSettingsPatchForModel(value);
+    const metaSettingsPatch = metaSettingsPatchForModel(value);
     setTranscriptionModel(value);
     try {
       if (value === "soniox-async") {
@@ -3330,6 +3354,8 @@ export default function Settings() {
         await updateSettings({ defaultSttService: "modulate_async" });
       } else if (value === "modulate-realtime") {
         await updateSettings({ defaultSttService: "modulate" });
+      } else if (metaSettingsPatch) {
+        await updateSettings(metaSettingsPatch);
       } else if (geminiSettingsPatch) {
         await updateSettings(geminiSettingsPatch);
       } else if (value === "mistral-async") {

README.md

@@ -213,6 +213,48 @@ fee, not a markup on the model's inference price.
 [See the OpenRouter model page.](https://openrouter.ai/microsoft/mai-transcribe-1.5)
 [See the OpenRouter fee details.](https://openrouter.ai/docs/faq)
 
+### Meta Muse Voice Transcribe
+
+Choose **Meta Muse Voice Transcribe Realtime** for live microphone transcription,
+or **Meta Muse Voice Transcribe Async** to transcribe after stopping. Both use
+`muse-voice-transcribe-1.0`, released by Meta on September 1, 2026, and the
+existing **Meta Model API** credential in Settings (`MODEL_API_KEY`). Model API
+access must be enabled for your Meta project; a Muse Spark entitlement alone
+does not prove Voice access.
+
+Realtime displays replaceable partials and injects only completed speech turns.
+It does not run local VAD or speaker detection. Async buffers one recording and
+submits one HTTP upload; Meta documents no separate async model or job-polling
+endpoint. File, YouTube and Meeting final transcription use that same upload
+route, with native speaker labels when requested.
+
+Limits: 60 minutes per realtime session; 10 minutes per upload. Uploads must fit
+32 MB including multipart overhead. Scriber prepares mono PCM16 WAV at 16 kHz
+when necessary and also accepts verified 24 kHz mono PCM16 WAV unchanged.
+There are 25 documented languages including German; use automatic language
+detection or a supported language hint. Custom vocabulary becomes Meta's
+`keywords`. Turn timestamps are approximate processed-audio boundaries, not
+word timestamps. The published rate is $0.18 per audio hour for both endpoints.
+
+Optional environment configuration (the UI stores credentials through the
+existing secret configuration path):
+
+```dotenv
+MODEL_API_KEY=your-meta-model-api-key
+SCRIBER_DEFAULT_STT=meta_stt
+# Use meta_stt_async above for transcription after stop.
+SCRIBER_META_STT_MODEL=muse-voice-transcribe-1.0
+```
+
+Unknown model IDs fail closed until their audio contract is verified. Cancellation
+never submits buffered audio. Failed uploads and interrupted streams are not
+automatically replayed. No local model download or new ML dependency is needed.
+
+[Meta announcement](https://research.meta.ai/blog/introducing-muse-voice-transcribe),
+[Voice guide and limits](https://dev.meta.ai/docs/speech-to-text),
+[HTTP API](https://dev.meta.ai/docs/api-reference/voice/transcribe),
+[WebSocket protocol](https://dev.meta.ai/docs/api-reference/voice/realtime).
+
 ### Modulate.AI multilingual transcription
 
 Modulate.AI is available for multilingual batch and realtime transcription.

docs/ARCHITECTURE.md

@@ -7,6 +7,44 @@ architecture notes and should be updated when ownership boundaries change.
 
 ## Runtime Overview
 
+### Meta Muse Voice Transcribe (verified 2026-09-02)
+
+`src/meta_stt.py` owns the two Voice API contracts for
+`muse-voice-transcribe-1.0`. `meta_stt` authenticates the WebSocket in its first
+JSON frame, sends paced raw mono PCM16, and uses provider-native ENDPOINTING.
+Interim hypotheses are replaceable `InterimTranscriptionFrame` previews, never
+durable text. Completed turns are keyed by `turnId`, emitted in speech-start
+order, and deduplicated. Neither `transcript.final` nor `speechEnd` commits a
+turn. Stop drains admitted audio, sends `endStream`, and waits for close 1000;
+errors or unfinished turns cannot become successful completion. The receiver
+belongs to an `AsyncTaskSupervisor`, and the application retains ownership of
+the shared HTTP session.
+Connection setup, audio sends and socket closure share one lock. Audio waits
+for handshake acknowledgement and rechecks shutdown state after waiting for
+startup or pacing; cancellation cannot leak a socket or send queued audio.
+
+`meta_stt_async` buffers one recording in the existing disk-backed PCM spool
+and submits one multipart HTTP request on normal stop. Cancel discards it.
+The same HTTP path serves File, YouTube and Meeting finals. It uses buffered
+JSON (`Accept: application/json`), not SSE or a speculative job-polling API.
+Both routes use `MODEL_API_KEY`; no model weights or additional dependencies
+are required. The immutable route binds the model, vocabulary, language,
+endpoint identity and exact audio-format capability. WAV pass-through also
+checks sample rate and channel count. `DIARIZATION` is requested only for
+speaker-enabled file/Meeting results; normal live dictation has no speakers.
+Native turns remain `provider_segment` evidence and never become fabricated
+word timestamps. Uploads are limited to ten minutes and 32 MB including
+multipart overhead; realtime sessions are limited to one hour. There is no
+automatic replay after failure, no session resumption and no automatic
+long-recording splitting.
+
+Protocol sources: [guide](https://dev.meta.ai/docs/speech-to-text),
+[upload](https://dev.meta.ai/docs/api-reference/voice/transcribe),
+[realtime](https://dev.meta.ai/docs/api-reference/voice/realtime),
+[schemas](https://dev.meta.ai/docs/api-reference/voice/schemas).
+
+### Application components
+
 Scriber is a hybrid desktop app:
 
 - Tauri 2 shell for installed Windows desktop runtime.

docs/ROADMAP_AND_KNOWN_ISSUES.md

@@ -7,6 +7,15 @@ It tracks current status only.
 
 ## Recently Completed
 
+Meta Muse Voice Transcribe integration (2026-09-02): separate Realtime and Async
+choices use the same `muse-voice-transcribe-1.0` model and Meta Model API key.
+Async means one non-blocking HTTP upload, not a separate Meta model or hosted
+background job. Current constraints: ten minutes per file/Meeting final,
+60 minutes per realtime session, no automatic splitting/resume, and no word
+timestamps. HTTP SSE is available from Meta but not used by Scriber; realtime
+previews use the WebSocket endpoint. Protocol tests are local; production
+account access and installed microphone smoke still require verification.
+
 Desktop runtime:
 
 - Tauri is the primary Windows desktop runtime.

docs/TESTING_AND_RELEASE.md

@@ -7,6 +7,19 @@ notes.
 
 ## Core Test Commands
 
+Muse Voice Transcribe regression coverage is in `tests/test_meta_stt.py`.
+It exercises real local aiohttp HTTP/WebSocket servers: multipart settings and
+auth, PCM framing, cumulative previews, overlapping turn completion, endStream
+drain, authentication rejection, disconnects, timeout, privacy-safe errors,
+no retry, cancel-without-upload, model/language validation and upload limits.
+Startup-race tests initialize the real Pipecat processor and use server-side
+barriers at both WebSocket upgrade and handshake acknowledgement. They check
+waiting audio and cancellation before client-session cleanup can hide leaks.
+Run it together with audio-preparation, route-artifact, config and pipeline-stop
+tests. Live Meta validation requires an explicitly configured `MODEL_API_KEY`
+with Voice access; local protocol tests do not prove account entitlement or
+recognition quality. No installer was built as part of this provider change.
+
 Run from repository root unless specified.
 
 Python (always through Scriber's project environment, never bare `python`):

src/api/upload_policy.py

@@ -62,6 +62,8 @@
     "openrouter_stt": (300 * 1024 * 1024, "300MB"),
     "modulate": (100 * 1024 * 1024, "100MB"),
     "modulate_async": (100 * 1024 * 1024, "100MB"),
+    "meta_stt": (32_000_000 - 65_536, "32MB including multipart overhead"),
+    "meta_stt_async": (32_000_000 - 65_536, "32MB including multipart overhead"),
 }
 
 

src/audio_prepare.py

@@ -301,6 +301,11 @@ def resolve_provider_audio_selection(
         else None
     )
     original = probe.audio_format if effective_limit is None or probe.byte_length <= effective_limit else None
+    if provider in {"meta_stt", "meta_stt_async"}:
+        if probe.duration_ms is not None and probe.duration_ms > 600_000:
+            raise ProviderAudioPreparationError("Meta STT recordings must not exceed 10 minutes.")
+        if probe.channels != 1 or probe.sample_rate not in (16_000, 24_000):
+            original = None
     selection = select_audio_input_format(
         capability,
         route_kind=ProviderAudioRouteKind.BATCH,
@@ -386,6 +391,12 @@ async def prepare_provider_audio_file(
             and frozen_selection.audio_format != probe.audio_format
         ):
             raise ProviderAudioPreparationError("Frozen pass-through format does not match the probed source.")
+        if (
+            provider in {"meta_stt", "meta_stt_async"}
+            and frozen_selection.mode == AudioSelectionMode.ORIGINAL_PASSTHROUGH
+            and (probe.channels != 1 or probe.sample_rate not in (16_000, 24_000))
+        ):
+            raise ProviderAudioPreparationError("Frozen Meta WAV must be mono PCM16 at 16 or 24 kHz.")
         selected = frozen_selection
 
     generated = selected.mode != AudioSelectionMode.ORIGINAL_PASSTHROUGH

src/config.py

@@ -295,6 +295,8 @@ class Config:
     DEFAULT_OPENROUTER_STT_MODEL = "microsoft/mai-transcribe-1.5"
     DEFAULT_GEMINI_STT_MODEL = "gemini-3.5-transcribe"
     DEFAULT_GEMINI_REALTIME_STT_MODEL = "gemini-3.5-transcribe-live"
+    DEFAULT_META_STT_MODEL = "muse-voice-transcribe-1.0"
+    META_STT_MODEL = os.getenv("SCRIBER_META_STT_MODEL", DEFAULT_META_STT_MODEL)
     _LEGACY_DEFAULT_GEMINI_STT_MODELS: ClassVar[set[str]] = {"gemini-2.5-flash"}
 
     # API Keys
@@ -409,6 +411,8 @@ class Config:
         "speechmatics_async": "SPEECHMATICS_API_KEY",
         "modulate": "MODULATE_API_KEY",
         "modulate_async": "MODULATE_API_KEY",
+        "meta_stt": "MODEL_API_KEY",
+        "meta_stt_async": "MODEL_API_KEY",
         "onnx_local": None,  # No API key needed for local models
     }
 
@@ -438,6 +442,8 @@ class Config:
         "speechmatics_async": "Speechmatics (Batch)",
         "modulate": "Modulate (Realtime Multilingual)",
         "modulate_async": "Modulate (Batch Multilingual)",
+        "meta_stt": "Meta Muse Voice Transcribe (Realtime)",
+        "meta_stt_async": "Meta Muse Voice Transcribe (Async)",
         "onnx_local": "Local (ONNX)",
     }
 
@@ -845,6 +851,8 @@ def configured(value: object, fallback: str) -> str:
             "soniox-async": configured(cls.SONIOX_ASYNC_MODEL, cls.DEFAULT_SONIOX_ASYNC_MODEL),
             "modulate-realtime": "velma-2-stt-streaming",
             "modulate-async": "velma-2-stt-batch",
+            "meta-realtime": cls.META_STT_MODEL,
+            "meta-async": cls.META_STT_MODEL,
             "gemini-stt": configured(cls.GEMINI_STT_MODEL, cls.DEFAULT_GEMINI_STT_MODEL),
             "gemini-realtime": configured(
                 cls.GEMINI_REALTIME_STT_MODEL,
@@ -1042,6 +1050,7 @@ def add(k, v):
         add("GROQ_API_KEY", cls.GROQ_API_KEY or "")
         add("SPEECHMATICS_API_KEY", cls.SPEECHMATICS_API_KEY or "")
         add("MODULATE_API_KEY", cls.MODULATE_API_KEY or "")
+        add("SCRIBER_META_STT_MODEL", cls.META_STT_MODEL)
 
         add("SCRIBER_HOTKEY", cls.HOTKEY)
         add("SCRIBER_POST_PROCESSING_HOTKEY", cls.POST_PROCESSING_HOTKEY)