beam_audio

beam_audio runs a voice turn (speech-to-text, agent, text-to-speech) and gives you the pieces for a WebSocket protocol. It does not include a WebSocket server.

beam_audio = { path = "../riverbeam/crates/beam_audio", features = ["audio-http-stt"] }

Feature	Effect
`audio-http-stt`	`HttpStt`
`audio-whisper`	Alias for `audio-http-stt`
`audio-decode`	WebM input reports `BEM-352` instead of `BEM-353`

Types

Item	Role
`run_audio_turn(audio, seq, session_id, stt, tts, invoker, sink)`	One turn
`SttBackend`, `TtsBackend`, `AudioInvoker`, `AudioSink`	The traits you implement or pick
`MockStt`, `MockTts`, `HttpStt`	Shipped backends
`AudioSession`, `SessionState`, `can_transition`, `AppendOutcome`	Session state and buffering
`ControlMessage`, `ControlType`, `ServerMessage`	WebSocket messages
`pcm_f32_to_wav_bytes`, `parse_wav_pcm_f32`, `resample_pcm`, `decode_to_pcm_16k`, `is_probable_webm`	Audio formats
`MAX_AUDIO_CHUNK_BYTES`, `MAX_AUDIO_BUFFER_BYTES`, `MIN_UTTERANCE_BYTES`, `MAX_MSG_PER_SEC`, `HEARTBEAT_INTERVAL_SEC`, `HEARTBEAT_TIMEOUT_SEC`	Limits

The protocol and the turn's message order are in Realtime audio.

Types​

Types