Overview

The Krisp API gives you Krisp's speech models as realtime WebSocket endpoints. You open one WebSocket connection per session, describe the session in a JSON message, then stream audio or text and receive results on the same connection as they are produced.

Endpoints

Region and latency

The Krisp API currently runs in the US region. This applies to all endpoints: Speech-to-Text, Text-to-Speech, Voice Translation and Voice Isolation. You can connect from anywhere, but clients outside the US see extra network latency, which grows with the distance from the US. For the lowest latency, connect from servers in or near the US.

Authentication

The STT, TTS and Voice Isolation endpoints authenticate with a session token that your backend generates from your Krisp API key. See Authentication for how to generate one; do not send your API key (master token) on the WebSocket. Voice Translation uses the same session tokens, with usage_type vt. Send the session token in the Authorization header of the WebSocket handshake, using the api-key scheme (case-insensitive, so Api-Key also works):

Authorization: api-key SESSION_TOKEN

If your client cannot set headers on a WebSocket handshake (for example, browsers), pass the same value as the authorization query parameter instead, URL-encoded:

wss://streaming.krisp.ai/stt?authorization=api-key%20SESSION_TOKEN

If authentication fails, the server sends an error message (see below) and closes the connection with close code 1008.

Session protocol

All endpoints follow the same flow:

1. Connect to wss://streaming.krisp.ai/<endpoint> with your session token.
2. Send the client hello. The first message must be a JSON text message that configures the session (each endpoint page documents its fields). It must arrive within 15 seconds of connecting.
3. Wait for the server hello. The server answers with either a ready message (session started) or an error message, after which it closes the connection.
4. Stream. After ready, send binary audio frames and/or JSON text messages, as described on the endpoint page. Results arrive on the same connection.
5. Close the WebSocket when you are done.

Each message you send after the hello must be at most 65,536 bytes (64 KB). Larger messages are rejected with the chunk_too_large error and the connection is closed.

ready: GatewayServerHello

The first response sent by the server to the client. Successful handshake response format.

Fields with a null value are omitted from the message. Some endpoints add their own fields to ready; see each endpoint page. Keep request_id: it identifies the session if you contact Krisp support.

{"type": "ready", "request_id": "3f2a9c1e-7b4d-4c2a-9e51-0d6f8a2b7c90", "metadata": {"call_id": "abc-123"}}

error: ErrorResponse

Error response schema following the standard format.

ErrorInfo

Error information structure.

After sending an error message, the server closes the WebSocket. The close reason is the same as error.message.

{
  "type": "error",
  "request_id": "3f2a9c1e-7b4d-4c2a-9e51-0d6f8a2b7c90",
  "error": {
    "code": "invalid_client_hello",
    "status_code": 400,
    "message": "Invalid ClientHello: sample_rate: Input should be less than or equal to 16000",
    "details": {
      "hint": "Check documentation for the correct ClientHello message format",
      "errors": [{"field": "sample_rate", "message": "Input should be less than or equal to 16000"}]
    }
  }
}

Common errors

These errors can occur on every endpoint. Endpoint-specific errors are listed on each endpoint page.

Changelog

• 2026-10-01: Voice Translation authentication now points to its own session token endpoint.
• 2026-10-01: Authentication now uses a session token; see Authentication.
• 2026-10-01: Page created: endpoints, authentication, session protocol, common errors.