超出任一限制将返回 413 Payload Too Large 或 422 Validation Error—— 参见 §05 错误码。

Exceeding any limit returns 413 Payload Too Large or 422 Validation Error — see §05 Error Codes.

说明：其他格式组合会被拒绝并立刻关闭（close code 1008）。 浏览器端可直接用 AudioContext({ sampleRate: 16000 })，由浏览器自动重采样到 16 kHz。

Any other format combination is rejected with an immediate close (close code 1008). Browsers can just use AudioContext({ sampleRate: 16000 }) and let the browser resample.

图中数值仅用于说明 P50/P99 的读法。实际句尾延迟以 revision_result.stats.endpoint_to_result_ms 为准。

The chart illustrates how to read P50/P99. Measure actual endpoint latency from revision_result.stats.endpoint_to_result_ms.

点击话筒开始录音；每说完一句话，右侧就会实时多出一段结果。再次点击话筒或发送 close 触发最终结果。需要上传文件请看 说话人日志 · 文件。

Click the mic to start recording; after each utterance ends, a new segment appears on the right. Click the mic again or send close to finalize. For file uploads, see Speaker Diarization · File.

先看下方消息序列 + 字段定义，再切到最下方语言 tab 复制可直接运行的代码片段。 不需要实时请看 说话人日志 · 文件模式。

Start by reading the message sequence and field definitions below, then scroll to the language tabs for copy-paste-ready snippets. Don't need real-time? See Speaker Diarization · File mode.

WebSocket 握手 → 发 start → 持续推 PCM 帧 → 按顺序应用 revision_result 的 speaker_merges 和 operations → revision 不连续时请求 snapshot → 发 close 得到最终快照。

WebSocket handshake → send start → stream PCM frames → apply each revision_result's speaker_merges and operations in order → request a snapshot after a revision gap → send close to receive the terminal snapshot.

长会议快照：服务端每 60 个 revision 或 5 分钟主动发送一次 snapshot_result。它是截至当前的完整状态，会随会议变长；前端不要自行设置 64 KiB 消息上限。发现 revision 缺口后只请求一次快照，在 snapshot_result 到达前暂停应用新增量。

Long-session snapshots: the server proactively emits a snapshot_result every 60 revisions or five minutes. It is the complete state so far and grows with the meeting; do not impose an application-level 64 KiB message limit. After a revision gap, request one snapshot and pause delta application until that snapshot_result arrives.

请求字段

Request fields

消息序列

Message sequence

响应样本

Response sample

// 先应用 speaker_merges，再应用 operations
{
  "type": "revision_result",
  "session_id": "6261ae2e072449438ee3d06f8e822ba1",
  "base_revision": 6,
  "revision": 7,
  "trigger": "utterance_end",
  "observed_through": 12.8,
  "speaker_merges": [{ "from_speaker_id": 3, "into_speaker_id": 1 }],
  "operations": [{
    "op": "upsert",
    "segment": { "segment_id": "seg-00000007", "start_time": 9.2, "end_time": 12.4, "speaker_id": 1, "text": "", "provisional": false }
  }],
  "stats": {
    "endpoint_to_result_ms": 386.4,
    "pipeline_ms": 214.7
  }
}

代码示例

Code snippets