Controller & Transport

Why controller decides transport

Providers expose both methods where available (stream? for WS, transcribe? for HTTP). The controller picks the active transport per session:

• transport: 'websocket' | 'http' | 'auto'
• Switch behavior without changing provider code

This keeps providers simple and lets apps tune transport per UX/cost needs.

WebSocket — silence policy

createTranscription({
  provider,
  recorder,
  transport: 'websocket',
  connection: { ws: { silencePolicy: 'keep' } }, // 'keep' | 'drop' | 'mute'
});

• keep (default): send all frames
• drop: send only during speech
• mute: keep cadence with zeroed frames in silence

Use drop to save bandwidth or mute when provider expects constant cadence.

HTTP — chunking & “segment‑only”

createTranscription({
  provider,
  recorder,
  transport: 'http',
  flushOnSegmentEnd: true,
  connection: { http: { chunking: { intervalMs: 0, overlapMs: 500, maxInFlight: 1 } } },
});

Chunking options:

• intervalMs: timer‑based flush (set to 0 for segment‑only)
• minDurationMs: minimum audio before a timer flush
• overlapMs: tail continuity between chunks
• maxInFlight: limit parallel HTTP flushes
• timeoutMs: per‑flush timeout

With flushOnSegmentEnd: true + intervalMs: 0 you get one request per phrase.

Retries (WS)

createTranscription({
  provider,
  recorder,
  transport: 'websocket',
  connection: { ws: { retry: { enabled: true, maxAttempts: 3, baseDelayMs: 300 } } },
});

On transient errors the controller reconnects with exponential backoff.

Force flush / endpoint

• HTTP: controller.forceEndpoint() triggers a flush of the current buffer.
• WS: if provider supports explicit endpointing, controller forwards forceEndpoint() to the stream.