Voice & meetings

Capture paths

1. Phone dial-in — Twilio + Gemini Live. Real-time speech-to-text, zero upload. See Voice dial-in.
2. File upload — mp3 / m4a / webm. Whisper transcribes. See Uploading meetings.
3. Paste transcript — skip transcription. Fastest path.

All three produce a Meeting row with a Transcript. Downstream is identical.

Transcription

Whisper (default)

• Cloud: OpenAI’s whisper-1 model. Requires OPENAI_API_KEY.
• Self-hosted: any Whisper-compatible endpoint via WHISPER_BASE_URL=http://host:9000. Run whisper.cpp or Faster-Whisper.

Domain-aware prompting

Before transcription, we build a domain prompt from:

• The active project’s name.
• Top-20 god-nodes from the project graph (technical terms, class names).
• Frequently-referenced nouns from previous transcripts in the same project.

This prompt is passed as the prompt field to Whisper. The difference on technical meetings is significant — “BAAgentService” stops coming back as “be agent service.”

See backend/src/services/transcription/domain-prompt.ts.

Gemini Live (voice dial-in only)

For real-time dial-in, we use Gemini Live which streams audio in and emits text + semantic events. Lower latency than Whisper; transcription quality is comparable for conversational speech.

The Meeting object

Stored in Postgres:

• id — stable.
• tenantId, projectId — scope.
• title — inferred from filename / Meet calendar / user input.
• source — upload | voice | paste.
• duration — seconds; null for paste.
• transcript — full text.
• speakers — if parseable from source.
• status — uploading | transcribing | ready | failed.
• audioUrl — optional; only when RETAIN_RECORDINGS=1.

Lifecycle

[source]
   │
   ▼
Meeting.status = transcribing
   │
   ▼  (Whisper / Gemini Live)
Transcript persisted
   │
   ▼
Meeting.status = ready
   │
   ▼  (if FEATURE_AUTO_BRIEF=1)
Brief drafting kicked off

Editing a transcript

Transcripts are versioned (Transcript.version). Edits produce a new version; old versions are retained for audit. The chief-of-staff uses the latest version on next brief draft.

Privacy flags

Flag	Default	Effect
RETAIN_RECORDINGS	0	Keep raw audio after transcription.
RETAIN_CALL_AUDIO	0	Keep Twilio call audio (post-stream).
WHISPER_BASE_URL	unset	Use local Whisper endpoint.
VOICE_CONSENT_GREETING	1	Prepend a recording-consent notice.

Limitations

• No speaker diarization on phone dial-in (Gemini Live limitation).
• No real-time transcript visible during a call (UI polls every 5s).
• Single-language per meeting; no auto-language-detect.

Typical latency

Path	End-to-end	Notes
Paste	<100 ms	Just a DB insert.
Upload 10-min	~30 s	Whisper inference + Drive download if relevant.
Upload 1-hour	~90 s	Linear-ish in audio length.
Voice dial-in	transcript ready within 5 s of hangup	Live streaming means text is ready when the call ends.