Realtime streaming

Stream PCM audio over WebSocket and receive partial and final results.

Connect to:

wss://go.monologue.to/v1/enterprise/streaming

Use the monologue.dictate.realtime.v1 protocol. Realtime sessions accept mono 16 kHz signed 16-bit little-endian PCM (pcm_s16le) and support smart or transcription as the source.

Authenticate

Non-browser clients can send X-API-Key or an Authorization: Bearer client token during the WebSocket handshake.

Browsers cannot set those headers. Mint an origin-bound client token, encode it as unpadded base64url, and send it as a credential subprotocol:

const credential = btoa(clientToken).replaceAll("+", "-").replaceAll("/", "_").replaceAll("=", "");

const socket = new WebSocket("wss://go.monologue.to/v1/enterprise/streaming", [
  "monologue.dictate.realtime.v1",
  `monologue.enterprise.bearer.${credential}`,
]);

Start the session

Send session.start before any audio. transcription_id must be nonblank.

{
  "type": "session.start",
  "protocol_version": 1,
  "transcription_id": "0198b68f-c6e1-7891-a341-803eb8fe421f",
  "source": "smart",
  "languages": ["en"],
  "platform": "web",
  "app_version": "1.0.0",
  "audio": {
    "encoding": "pcm_s16le",
    "sample_rate_hz": 16000,
    "channels": 1
  },
  "dictionary": "Monologue, Avery",
  "context": "<app_name>Slack</app_name>"
}

Wait for session.ready, then send raw PCM bytes as binary WebSocket frames. Do not wrap them in JSON or base64. When audio is complete, send:

{
  "type": "input.commit",
  "extra_instructions": "Write a concise Slack message."
}

Server events

Event	Meaning
`session.ready`	The server is ready for audio.
`transcript.partial`	Unstable partial text for live UI.
`transcript.segment`	A stable segment and accumulated text.
`transcript.final`	Final raw transcript after commit.
`result.final`	Final `text`, `raw_text`, output mode, and timing data.
`session.completed`	The session finished successfully or was cancelled.
`session.error`	The session failed; inspect `code` and `retry_via_async`.

You may send context.update before commit to replace context, dictionary, or extra instructions. Send session.cancel to stop without a final result. Audio frames are rejected after input.commit.