Streaming transcription

Open a WebSocket to wss://api.speechrouter.ai/v1/listen, send PCM audio frames, receive normalized transcript events. Same contract for all 12 providers.

Query parameters

Param	Default	Meaning
model	(required)	Model slug, e.g. deepgram/nova-3
fallbacks	—	Comma-separated failover lane, e.g. soniox/stt-rt-v5,openai/gpt-4o-transcribe
encoding	linear16	PCM encoding of the audio you send
sample_rate	16000	Sample rate in Hz
channels	1	Channel count
language	auto	BCP-47 hint where the model supports it
interim_results	true	Emit non-final hypotheses
diarization	false	Speaker labels (see per-model support)
keyterms	—	Comma-separated bias terms
provider_params	—	URL-encoded JSON object, passed through raw
include_raw	false	Attach untouched provider payload to each transcript

Auth goes in the Authorization header or the ['bearer', key] subprotocol — see Authentication.

Sending audio

• Binary frames are audio, in the encoding/rate you declared. Chunks of 20–250 ms work well everywhere (providers with stricter pacing rules are re-paced by the gateway).
• Text frames are JSON control messages:

{ "type": "finalize" }    // flush: force finals for audio sent so far, then done
{ "type": "keepalive" }   // keep an idle connection open

The event stream

{ "type": "session.open", "session_id": "…", "model": "deepgram/nova-3" }
{ "type": "speech_started", "at": 0.48 }
{ "type": "transcript", "is_final": false, "text": "speech router is", "start": 0.5 }
{ "type": "transcript", "is_final": true,  "text": "SpeechRouter is live.",
  "words": [{ "w": "SpeechRouter", "start": 0.5, "end": 0.9, "speaker": 0 }],
  "start": 0.5, "end": 1.9 }
{ "type": "utterance_end", "at": 1.9 }
{ "type": "done", "usage": { "audio_seconds": 12.4 } }

Full schemas: WebSocket events.

• transcript.words[] carries per-word timing (audio-time seconds from session start), confidence where available, and integer speaker labels when diarization is on.
• utterance_end is the endpointing signal — the “respond now” trigger for voice agents.
• If a provider dies mid-session you’ll see { "type": "provider_switched", … } and the stream continues on the next model in your lane — failover explains the guarantees.

Ending a session

Send { "type": "finalize" } and keep reading until done, or just close the socket. Sessions with no audio for 60 seconds are closed as audio_timeout (send keepalive to hold an idle line open).