pyannoteAI WebSocket Gateway API

asyncapi: '3.0.0'
info:
  title: pyannoteAI WebSocket Gateway API
  version: '0.0.1-beta'
  description: Streaming WebSocket gateway for real-time speaker diarization.
  termsOfService: https://www.pyannote.ai/terms-of-use
  contact:
    name: pyannoteAI Support
    url: https://pyannote.ai/
    email: support@pyannote.ai
  externalDocs:
    description: Find more info here
    url: https://docs.pyannote.ai
channels:
  root:
    address: <generated url from /v1/live endpoint>
    bindings:
      ws:
        query:
          type: object
          description: >
            The token is embedded in the WebSocket URL returned by the
            POST /v1/live REST endpoint. It is single-use and grants access
            only to this stream session — it carries no team credentials or API key.
          properties:
            token:
              type: string
              description: Single-use temporary token generated when creating a stream session via POST /v1/live.
    messages:
      binaryAudioChunk:
        $ref: '#/components/messages/binaryAudioChunk'
      endOfStream:
        $ref: '#/components/messages/endOfStream'
      diarizationSpeakerStart:
        $ref: '#/components/messages/diarizationSpeakerStart'
      diarizationSpeakerEnd:
        $ref: '#/components/messages/diarizationSpeakerEnd'
      error:
        $ref: '#/components/messages/error'
operations:
  sendAudioChunk:
    action: receive
    summary: Stream a raw PCM audio chunk to the server.
    description: >
      Send 100 ms of raw PCM audio as a binary WebSocket frame. Audio must be
      PCM float 32-bit little-endian (pcm_f32le), 16 kHz, mono. Do not include
      any file headers (e.g. WAV/RIFF headers). The server enforces a maximum
      5-second buffer; sending audio faster than real-time will close the
      connection.
    channel:
      $ref: '#/channels/root'
    messages:
      - $ref: '#/channels/root/messages/binaryAudioChunk'
  sendEndOfStream:
    action: receive
    summary: Signal that no more audio will be sent.
    description: >
      Send this JSON text frame when you have finished streaming audio. The
      server will finalize diarization, emit any remaining events, and close
      the connection with code 1000 (normal closure). Do not send further audio
      frames after this message. Using this is recommended over abruptly closing
      the socket, which may cause final events to be lost.
    channel:
      $ref: '#/channels/root'
    messages:
      - $ref: '#/channels/root/messages/endOfStream'
  receiveSpeakerStart:
    action: send
    summary: A speaker begins a turn.
    channel:
      $ref: '#/channels/root'
    messages:
      - $ref: '#/channels/root/messages/diarizationSpeakerStart'
  receiveSpeakerEnd:
    action: send
    summary: A speaker's turn ends.
    channel:
      $ref: '#/channels/root'
    messages:
      - $ref: '#/channels/root/messages/diarizationSpeakerEnd'
  receiveError:
    action: send
    summary: The server encountered a problem processing a frame.
    channel:
      $ref: '#/channels/root'
    messages:
      - $ref: '#/channels/root/messages/error'
components:
  messages:
    binaryAudioChunk:
      name: binaryAudioChunk
      summary: Raw PCM audio chunk (binary frame).
      description: >
        100 ms of PCM float 32-bit little-endian audio at 16 kHz mono.
        Each chunk contains 1600 samples (6400 bytes). Send raw bytes only —
        no container or file headers.
      contentType: application/octet-stream
      payload:
        $ref: '#/components/schemas/binaryAudioChunk'
    endOfStream:
      name: endOfStream
      summary: End-of-stream signal.
      description: >
        Sent by the client to indicate no further audio frames will follow.
        Optional but strongly recommended to ensure all diarization events are
        flushed before the connection closes.
      contentType: application/json
      payload:
        $ref: '#/components/schemas/endOfStream'
    diarizationSpeakerStart:
      name: diarizationSpeakerStart
      summary: Emitted when a speaker begins a turn.
      contentType: application/json
      payload:
        $ref: '#/components/schemas/diarizationSpeakerStart'
    diarizationSpeakerEnd:
      name: diarizationSpeakerEnd
      summary: Emitted when a speaker's turn ends.
      contentType: application/json
      payload:
        $ref: '#/components/schemas/diarizationSpeakerEnd'
    error:
      name: error
      summary: Emitted when the server encounters a problem processing a frame.
      contentType: application/json
      payload:
        $ref: '#/components/schemas/error'
  schemas:
    binaryAudioChunk:
      type: string
      format: binary
      description: >
        Raw binary audio data. PCM float 32-bit little-endian (pcm_f32le),
        16 kHz sample rate, mono channel, 100 ms per chunk (1600 samples / 6400 bytes).
      examples: 
        - <binary pcm_f32le 100ms audio chunk>
    endOfStream:
      type: object
      required:
        - type
      properties:
        type:
          type: string
          const: end_of_stream
      examples:
        - type: end_of_stream
    diarizationSpeakerStart:
      type: object
      required:
        - type
        - data
      properties:
        type:
          type: string
          const: diarization_speaker_start
        data:
          type: object
          required:
            - timestamp
            - speaker
          properties:
            timestamp:
              type: number
              description: Time in seconds from the start of the stream when the speaker turn began.
              example: 1.24
            speaker:
              type: string
              description: Stable label for this speaker within the session. Up to 8 speakers are tracked simultaneously.
              example: SPEAKER_00
      examples:
        - type: diarization_speaker_start
          data:
            timestamp: 1.24
            speaker: SPEAKER_00
    diarizationSpeakerEnd:
      type: object
      required:
        - type
        - data
      properties:
        type:
          type: string
          const: diarization_speaker_end
        data:
          type: object
          required:
            - timestamp
            - speaker
          properties:
            timestamp:
              type: number
              description: Time in seconds from the start of the stream when the speaker turn ended.
              example: 3.86
            speaker:
              type: string
              description: Stable label for this speaker within the session.
              example: SPEAKER_00
      examples:
        - type: diarization_speaker_end
          data:
            timestamp: 3.863
            speaker: SPEAKER_00
    error:
      type: object
      required:
        - type
        - message
      properties:
        type:
          type: string
          const: error
        message:
          type: string
          description: Human-readable description of the error.
          example: Invalid chunk size
      examples:
        - type: error
          message: Invalid chunk size