For ongoing conversations or livestreams, you can have audio transcribed in real time. To do this, you open a session through the API and then connect over WebSocket. For finished recordings, use file transcription instead, see Transcribe audio files.
Is live available?
Live transcription is only available if GET /v1/models lists an audio model with the live mode under transcription. If there's no such model in your list, live isn't usable for you right now. File transcription keeps working regardless.
The prices per minute are listed with each model under pricing.variants.
Start a session
curl https://api.flowmodels.one/v1/audio/transcription-sessions \
-H "Authorization: Bearer $FLOWMODELS_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: conversation-2026-1042" \
-d '{"model": "openai/gpt-live-transcribe", "source": "push", "language": "en"}'
sourceispushfor audio from your application orurlfor a public audio stream, whose address you also send inurl.- By default, a session runs for at most two hours, and with
max_secondsfor up to eight hours. - At most three sessions can run at the same time per team.
The response contains id, websocket_url, ticket and ticket_expires_at. The ticket is valid for 60 seconds and is used up when you connect. It doesn't belong in the URL, you pass it as the WebSocket subprotocol flowai-ticket.TICKET. Never send your API key over the WebSocket.
Send audio
With source: "push", you send binary frames in PCM16 little-endian, mono, 24 kHz, at playback speed and with at most 48,000 bytes per frame. With source: "url", the server fetches the audio itself and you only receive events. Private network targets, URLs with credentials and playlists in HLS format aren't allowed.
{"type": "session.stop"} ends the recording. Pending results are still processed after that.
Receive results
These events, among others, come over the WebSocket:
transcript.partial: Interim state. Replace the previous draft with the samesection_id.transcript.final: Confirmed text with a runningsequence.connection.interrupted: The connection to the provider was interrupted.usage: Usage so far.
Remember the last confirmed sequence number. If the connection drops, with source: "push" you can reconnect within 30 seconds using a new ticket, which you get from POST /v1/audio/transcription-sessions/{id}/connection-ticket. You can catch up on missed sections with GET /v1/audio/transcription-sessions/{id}?after_sequence=N.
Costs and retention
Live sessions are billed by the minute. The price depends on the model and additional options, and silence can cost money too. Before the start, we initially reserve 60 seconds and keep extending the reservation as long as audio is being transmitted. Credit, key revocations and team limits also apply during the session. What was actually billed is shown in usage.cost.
You can fetch finished results for 24 hours after completion. You can delete a completed session with DELETE /v1/audio/transcription-sessions/{id}.
All fields and the complete flow example are in the documentation at flowmodels.one/en/docs.