Voice

Pro tier feature

Requires a Pro or higher licence key.

Bee Flow supports natural full-duplex voice conversations. You speak; the assistant speaks back; either side can interrupt.

Two modes

Mode	UX
Push-to-talk	Click & hold the microphone in the chat composer. Release to send. The transcript is submitted as the next message.
Voice call	Click Voice call in the chat header. The microphone stays open. The assistant detects when you stop and replies; you can interrupt mid-reply.

Voice call is the more interesting mode — most users land here after trying push-to-talk.

Voice call state machine

       ┌──────────┐
       │  IDLE    │  (after clicking "Voice call", before first speech)
       └────┬─────┘
            │ speech detected (energy > threshold)
            ▼
       ┌──────────┐
       │LISTENING │  (recording, VAD watching)
       └────┬─────┘
            │ silence > 900 ms, utterance ≥ 400 ms
            ▼
       ┌──────────┐
       │THINKING  │  (STT + agent + first TTS chunk)
       └────┬─────┘
            │ first audio chunk arrives
            ▼
       ┌──────────┐
       │SPEAKING  │  (playing TTS audio)
       └────┬─────┘
            │ done OR user interrupts (energy spike)
            ▼
       (back to IDLE / LISTENING)

Voice Activity Detection (VAD)

The VAD is energy-based, with hysteresis to prevent oscillation:

Constant	Default	Purpose
VAD_SPEECH_RMS	0.018	Start-of-speech threshold
VAD_SILENCE_RMS	0.012	End-of-speech threshold (lower = hysteresis)
VAD_SILENCE_MS	900	Trailing silence to end a turn
VAD_MIN_UTTERANCE_MS	400	Discard clips shorter than this (eats stray clicks)
VAD_MAX_UTTERANCE_MS	30 000	Hard cap at 30 seconds
VAD_FFT_SIZE	1024	Audio analysis block size
VAD_POLL_MS	50	Evaluate energy 20× per second

These tune well for typical office environments. In noisier rooms, raise VAD_SPEECH_RMS. With high-quality directional mics, lower it.

Barge-in

While the assistant speaks, the mic stays open. If the user starts speaking (energy spike), playback is interrupted immediately and a new turn begins. This simulates natural phone-conversation flow.

Provider routing

Direction	Provider	Notes
Speech → text (STT)	Voxtral (default)	Multilingual. Returns detected language with each turn.
	Deepgram (DEEPGRAM_API_KEY)	Alternative if you don't want Voxtral.
	Mistral Transcription (MISTRAL_TRANSCRIPTION_KEY)	EU-region option.
Text → speech (TTS)	Voxtral (default)	Streams chunks for low first-byte latency.
	ElevenLabs (ELEVENLABS_API_KEY)	Higher voice fidelity, more languages.
Local STT	Whisper (OLLAMA_BASE_URL + a Whisper model)	Fully offline option.

Switch provider in Settings → Organisation → Voice. Per-user voice (e.g. one Talk room uses a Dutch voice, another an English voice) is on the roadmap.

Sticky language

The first turn's detected language is cached as stickyLanguageRef for the session. Short utterances ("yes", "no") are notoriously hard to language-detect; sticking to the previous language prevents jarring mid-call language flips.

Browser support

Requires getUserMedia (WebRTC) and the Web Audio API:

Browser	Support
Chrome 80+	✅
Edge 80+	✅
Firefox 75+	✅
Safari 14+	✅
Mobile Safari (iOS 14+)	✅
Mobile Chrome (Android)	✅

The first time the user clicks Voice call, the browser asks for microphone permission. If denied, the button shows a permission-error tooltip.

Latency

Typical end-to-end latency from end-of-utterance to first audio chunk:

Layer	Time
STT	200–400 ms
LLM first token	300–800 ms (model-dependent)
TTS first chunk	150–300 ms
Total	0.7 – 1.5 s

The THINKING state shows a spinning bee animation so the user has feedback during the gap.

Privacy

The Privacy Shield runs on the transcribed text before it reaches the model — exactly the same as text chat. Voice inputs are not retained beyond session lifetime unless you explicitly enable conversation history for voice calls (off by default).

Audio bytes never hit the LLM provider — only transcripts do.

Server endpoint

Internally the SPA POSTs to /api/voice/turn with the recorded blob and recent message history. The server orchestrates STT → guardrails → agent → TTS and streams back.

Where to next

• Privacy shield — what's redacted from voice transcripts.
• API → Streaming (SSE) — same event protocol used for voice TTS chunks.