Alibaba Cloud Model Studio:WebSocket API for Qwen-TTS real-time synthesis

Service endpoint

Use the following URL. Set the model with the model query parameter:

Request headers

Set the following request headers:

Interaction flow

For detailed descriptions of client events and server events, see Client events and Server events.

The API supports two modes:

• ServerCommit mode: The server determines text segmentation and synthesis timing automatically.

• Commit mode: The client controls when to submit each text segment. Call input_text_buffer.commit explicitly to trigger synthesis.

Mode details

• In ServerCommit mode, call input_text_buffer.append to buffer text. The system determines synthesis timing automatically.

• Calling input_text_buffer.commit in ServerCommit mode synthesizes the current buffer immediately. The session then resumes ServerCommit mode.

• In Commit mode, input_text_buffer.append alone does not trigger synthesis. Call input_text_buffer.commit to start synthesis.

Key steps:

1. Connection: The client opens a WebSocket connection. The server returns session.created when ready.

2. Session configuration: The client sends session.update to configure voice, format, and mode.

3. Text input: The client sends input_text_buffer.append to add text to the buffer.

4. Synthesis trigger:

   • In ServerCommit mode, synthesis starts automatically. The client can call input_text_buffer.commit to force immediate synthesis.

   • In Commit mode, only input_text_buffer.commit triggers synthesis.

5. Audio generation: The server sends response.created, then streams base64-encoded audio as response.audio.delta events until response.audio.done.

6. Session termination: The client sends session.finish to release resources. The server returns session.finished and closes the connection.

After connection, the server returns this session.created event:

{
    "event_id": "event_xxx",
    "type": "session.created",
    "session": {
        "object": "realtime.session",
        "mode": "server_commit",
        "model": "qwen3-tts-flash-realtime",
        "voice": "Cherry",
        "response_format": "pcm",
        "sample_rate": 24000,
        "id": "sess_xxx"
    }
}