feat: integrate Meta Muse Voice Transcribe realtime and async

[MyButtermilk]

Summary

Corrects the earlier model identification: Meta released Muse Voice Transcribe on September 1, 2026. The Omnilingual ASR preparation commit is reverted; the net change integrates muse-voice-transcribe-1.0 instead.

• Add separate Realtime (meta_stt) and Async (meta_stt_async) choices in backend and frontend, reusing the existing Meta Model API credential (MODEL_API_KEY).
• Realtime: documented WebSocket handshake authentication, paced mono PCM16, replaceable partials, ENDPOINTING speechComplete finals, overlapping-turn ordering, deduplication, explicit endStream and terminal drain. No local VAD or live diarization.
• Async: one buffered HTTP multipart upload after normal stop; cancellation never submits the recording. File/YouTube/Meeting final transcription uses the same endpoint, with native diarization when requested. Async is a Scriber workflow around the same model, not a separate Meta model or an invented polling API.
• Bind model, language, vocabulary, endpoint and audio-format evidence to existing frozen routes. Validate WAV rate/channels, size and duration, preserve valid originals, normalize turn-level evidence without inventing word timestamps.
• Add bounded error handling, no automatic replay, shared transport ownership, runtime import coverage, localized settings, pricing/limit notes and configuration examples.

Verified primary sources

Research used Exa to retrieve current official Meta pages. Context7 was also queried, but did not provide applicable Voice API documentation; unrelated Responses API background-job documentation was not used.

• https://research.meta.ai/blog/introducing-muse-voice-transcribe
• https://dev.meta.ai/docs/speech-to-text
• https://dev.meta.ai/docs/api-reference/voice/transcribe
• https://dev.meta.ai/docs/api-reference/voice/realtime
• https://dev.meta.ai/docs/api-reference/voice/schemas

Both transports use muse-voice-transcribe-1.0. HTTP recordings: at most 10 minutes, 32 MB including multipart overhead, mono PCM16 WAV at 16/24 kHz. Realtime: raw PCM16 at 16/24 kHz, at most 60 minutes per session. Published pricing: $0.18/audio hour. No word timestamps or automatic long-recording splitting/resume.

Validation

• 360 focused backend regression tests passed, including local HTTP/WebSocket protocol tests, cancellation/no-retry, frozen-route execution, audio preparation, config, provider dependencies, pipeline stop, transcript artifacts and existing Modulate paths.
• Repository-wide Ruff lint and formatting passed (455 Python files).
• Scoped mypy passed (68 source files).
• Frontend app/test typechecks and repository ESLint passed with Node 26.5.0.
• 113 frontend library tests and 74 component tests passed.
• Production frontend build passed (existing large-chunk warning).
• git diff --check passed.

Remaining verification

No live Meta recognition/account-entitlement test was performed: no Meta credential is configured in this isolated worktree. Installed microphone smoke and installer/release validation were not run. No installer was built or release published. The original working checkout and its unrelated changes were left untouched.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f3da9c9d95

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".