Message Encoding - DeepL Documentation

WebSocket messages in a Voice API session can be encoded in two formats, chosen with the message_format option when you request the session. Start with JSON for the best developer experience, and switch to MessagePack if you need better performance.

Format	Frame type	Binary data	Trade-off
JSON (default)	TEXT	base64-encoded strings	Human-readable, easy to debug
MessagePack	BINARY	raw binary	Roughly 25-30% less bandwidth, 2x-4x faster encoding/decoding

Send JSON messages as TEXT frames and MessagePack messages as BINARY frames. Sending the wrong frame type results in connection errors.

MessagePack messages must be encoded as maps with string keys, not as arrays. The structure must match the JSON schema exactly, with all field names preserved (for example, {"source_media_chunk": {"data": <binary>}}). Some MessagePack libraries default to array encoding for performance, so check your library’s configuration. The following example sends the same audio chunk in both encodings:

// Raw binary audio data
const audioData = getAudioChunk();

// Base64 encode the audio data
const base64Audio = btoa(audioData);

const message = {
  source_media_chunk: {
    data: base64Audio
  }
};

// Send as TEXT frame
websocket.send(JSON.stringify(message));

import { pack } from 'msgpackr';

// Raw binary audio data
const audioData = getAudioChunk();

const message = {
  source_media_chunk: {
    data: audioData  // No base64 encoding needed
  }
};

// Send as BINARY frame
websocket.send(pack(message));

For the full list of message types exchanged during a session, see the WebSocket Streaming reference. For how those messages fit into the session lifecycle, see Understanding Voice Sessions.

Understanding Voice Sessions Supported Languages, Formats, and Limits

⌘I