FluidMeet: multilingual transcription with auto language detection (any language, incl. mixed Chinese + English)

[ericrcwu001]

Problem

FluidMeet can only transcribe meetings in English and the European languages Parakeet v3 supports. Many of my meetings are mostly Mandarin Chinese with English mixed in, sometimes within one sentence. FluidMeet can't transcribe that today:

• Both meeting backends are limited to Parakeet. MeetingParakeetNemotronBackend accepts only parakeetV3LanguageIDs (European languages), and MeetingLegacyCompatibilityBackend accepts only en. Chinese, Japanese, Korean, Arabic, Hindi and others are not supported.
• Each session has exactly one languageCode, which defaults to en with parakeet-tdt-v2. Nothing handles a meeting that switches between languages.
• The language has to be picked before recording, and a finished meeting can't be redone in another language ([BUG] Language not detected correctly in meeting transcript and cannot be changed #1033).

The dictation side already has multilingual models: Qwen3 ASR 1.7B, which as of v1.6.13-beta.1 auto-detects 30 languages including Chinese, plus Whisper, Cohere and Apple Speech. FluidMeet just can't use any of them.

This is the one thing keeping me on Granola for meetings. If FluidMeet handled mixed Chinese + English, I'd switch to it, since local, on-device meeting notes are exactly what I want.

Proposed solution

Let FluidMeet use a multilingual ASR model with automatic language detection, so it works for basically any language and for code-switched meetings (e.g. Chinese + English):

1. A multilingual meeting backend. Pair a multilingual ASR (Qwen3 ASR, Whisper large-v3/turbo, or similar) with the existing Nemotron diarization. Diarization works on the audio itself, so the speaker side should mostly carry over; only the ASR stage needs swapping.
2. "Auto" as a meeting language, and the default when a multilingual model is installed, with detection per segment or per speaker turn instead of one language locked for the whole session.
3. Mixed-language output. Keep each segment in its spoken language (Chinese stays in Chinese characters, English stays English), with no forced translation and no mangling into one script.
4. Store the detected language per segment in session.json, so retranscription ([BUG] Language not detected correctly in meeting transcript and cannot be changed #1033) and summaries can use it.

Alternatives considered

• Setting a single meeting language manually: this breaks for code-switched meetings, and Chinese isn't selectable for meetings at all.
• Exporting the audio and running it through a dictation model or an external tool: this loses diarization and the meeting UI.

Related: #1033 (retranscribe in another language), #506 (multiple active languages for multilingual models), #359 (SenseVoice).