声纹向量库

// 512 维声纹向量 — 紧凑 base64 编码,载荷约 684 字节
{
  "embedding": "AgUEAQQF/vwB+wTu/fwOAwD6APcHBwX/CQgCAvv+BgMICPz/...AAQ=",
  "dim": 512,
  "dtype": "int8",
  "encoding": "base64",
  "duration_seconds": 3.420
}

// 客户端解码(若需要 float 数组做点积):
//   bytes = base64.b64decode(embedding)            # 512 个有符号字节
//   vec   = np.frombuffer(bytes, dtype=np.int8)/127.0  # → float32, ≈ unit norm

// 同一个数组里可以混合两种条目类型
[
  // type=embedding —— 直接传预计算向量(/v1/voiceprint/extract 返回的 base64 字符串原样塞进来)
  {
    "id": "alice",
    "type": "embedding",
    "embedding": "AgUEAQQF/vwB+wTu/fwOAwD6APcHBwX/...AAQ="
  },
  // type=audio —— 现场上传音频,服务端自动 /v1/voiceprint/extract 后再比对
  {
    "id": "bob",
    "type": "audio",
    "audio_ref": "r0"     // 对应 multipart 字段 lib_audio_r0
  }
  // 兼容性:embedding 也可以传 [float32, ...] 数组(老格式),服务端自动识别。
]

{
  "threshold": 0.372,
  "top_k": null,
  "best_match": { "id": "alice", "score": 0.847, "is_same_speaker": true },
  "matches": [
    { "id": "alice", "score": 0.847, "is_same_speaker": true  },
    { "id": "bob",   "score": 0.183, "is_same_speaker": false },
    { "id": "carol", "score": 0.124, "is_same_speaker": false }
  ]
}