SpeechRelay

SpeechRelay is a streaming speech-to-text API: audio goes in over a WebSocket connection and transcript events come back over the same connection.

What it does

You open one bidirectional WebSocket stream per session, send audio as it is captured, and receive transcript events as they are produced. Two modes are offered: mt_asr transcribes several talkers in one conversation and tells them apart, and ts_asr follows one enrolled target speaker.

The service exposes /health, /ready, /metrics, /v1/metadata, session creation, the bidirectional WebSocket stream, and a whole-file compatibility endpoint.

Getting a key

Authentication comes in two forms. An operator can configure one shared credential for a whole fleet, delivered to the worker as a credential file and rotatable without causing authorization failures. Issued keys are the per-customer form: each one carries scopes, an optional expiry and the project it is attributed to, and each can be revoked. Revocation is not instant — it takes effect within the verification cache window plus storage propagation.

When a credential is configured, every route except /health and /ready requires it. The authentication page describes how the credential is configured, checked and rotated.

In this documentation

• Quickstart — the shortest path to a first transcript.
• Authentication — how the credential is configured, checked and rotated, and how failures surface.
• Streaming — the production create-and-stream path and the stream protocol.
• Resume — reattaching a dropped client to the same worker instance.
• Batch — whole-file compatibility for a WAV body in one request.
• Ordering, bounds and errors — delivery semantics, retryability and limits.
• Limits — the bounds a session is admitted and held under, and what the service does when one is reached.