Live transcription over WebRTC

Microphone audio streams straight to the OpenAI Realtime API through a peer connection. The server mints a short-lived client secret; the API key never reaches the browser. Every handshake stage is timed so time-to-first-word can be compared across settings.

gpt-live-transcribe · POST /v1/realtime/client_secrets · POST /v1/realtime/calls · dc oai-events

idle

——time to first word

Session settings

Baked into the ephemeral client secret at mint time — each run reconnects with exactly these values.

Model

Delay

Lower emits partial text sooner; higher gives more context.

Noise reduction

Languages

Englishen

Sent as ["en"]

Capture

Microphone

No input devices found yet.

Connection

Transport

Capture starts the instant you press Start; everything said during setup is buffered and flushed at session.created. Speak immediately — that is the point.

API region

The default endpoint. Always available. Applied to both the token mint and the SDP connection.

Start mode

Warm removes ~400 ms from Start, but changes what you measure — keep cold and warm runs apart when reading medians.

Press Start, allow the microphone, then speak.

Streaming deltas are grey and italic. An utterance is finalised only when a commit is sent — this model has no server-side turn detection, so the app commits after 1.5 s of delta silence.