Protocol

All control messages are JSON over WebSocket. Audio is binary PCM.

Client → Server

Start session:

{
  "type": "start",
  "sample_rate": 16000,
  "save_audio": true,
  "format": "wav",
  "target_language": "es"
}

• save_audio: bool, save recording on clean
• format: wav | ogg
• target_language: en|es|fr|de|pt — translated if model supports, otherwise ignored

Audio chunks: Binary PCM 16-bit mono @ 16kHz (only speech frames passed through VAD)

Stop / Clean:

{"type": "stop"}
{"type": "clean"}

clean = stop + save file to records/.

Server → Client

Transcription:

{"type": "transcription", "text": "..."}

Also accepted by client: partial, final.

Status / Error / Model info:

{"type": "status", "message": "ready"}
{"type": "error", "message": "..."}
{"type": "model_info", "model": "Canary180M", "languages": ["en","es","fr","de","pt"]}

model_info is sent on connect. Client uses it to show active model and enable only supported languages in Settings.

Example flow

Client                                    Server
  | -- {"type":"start", ...} -------------> |
  | -- [binary PCM chunk] ----------------> |
  | <-- {"type":"transcription","text":"hola"} |
  | -- [binary PCM chunk] ----------------> |
  | <-- {"type":"transcription","text":"hola mundo"} |
  | -- {"type":"clean"} ------------------> |
  | <-- {"type":"status","message":"saved records/..wav"} |