SpeechRelay

SpeechRelay is a streaming speech-to-text API: you send audio over a WebSocket connection and transcript events come back on it.

Try the playground #

The free playground is the way to try SpeechRelay today. It runs mt_asr with up to eight speakers per session, four at a time, and has limited availability. You sign in with Google or GitHub; each account has a limited amount of free use, and signing in again does not add more.

Integrate with the API #

API access is not open yet. It is opening gradually; join the waitlist. These pages document the protocol for callers who already have an API key; start with the Quickstart. ts_asr is documented here but is not available in the playground.

Terms used on these pages #

• Session — one transcription, from the session.create message to its final event. Each session has its own WebSocket connection.
• mt_asr — multi-talker mode. It transcribes a conversation between several people and labels each speaker.
• ts_asr — target-speaker mode. It follows one person, whose voice you send first as a short enrollment clip.
• dicon.v1 — the WebSocket subprotocol name. Every stream connection offers it.
• Partial and final — a transcript.partial adds words while someone is talking; a transcript.final carries that speaker's complete text. Streaming explains how to read them.

In this documentation #

• Quickstart — stream a WAV file and print its transcript.
• Authentication — send your key, and tell a rejected key from one without permission.
• Streaming — every message on the stream, and how a stream ends.
• Resume — reattach after a dropped connection, where the service enables it.
• Batch — send a whole WAV file in one request.
• Ordering, bounds and errors — the error format, every code, and what to retry.
• Limits — session bounds and quotas.
• Service endpoints — health and readiness probes.
• Pricing and quotas — the planned API price and the playground.
• Data retention — what is kept, for how long, and what is not stored.