Skip to content
Tutorials3 min read

說話人分離標註

構建一個帶有音訊波形、時間戳標記和說話人標籤分配的說話人識別任務。

Potato Team

說話人分離回答了"誰在什麼時候說話？"這個問題。本教程涵蓋構建標註說話人輪次的介面、糾正自動分離結果以及處理多說話人對話。

什麼是說話人分離？

說話人分離將音訊分割為說話人同質的區域。應用場景包括：

  • 會議轉錄
  • 客服中心分析
  • 播客製作
  • 訪談處理
  • 法庭/法律錄音

基本分離設定

yaml
annotation_task_name: "Speaker Diarization"
 
data_files:
  - "data/conversations.json"
 
annotation_schemes:
  - annotation_type: audio_annotation
    name: speakers
    description: "Mark when each speaker talks"
    labels:
      - name: Speaker 1
        color: "#FF6B6B"
        keyboard_shortcut: "1"
      - name: Speaker 2
        color: "#4ECDC4"
        keyboard_shortcut: "2"
      - name: Speaker 3
        color: "#45B7D1"
        keyboard_shortcut: "3"
      - name: Overlap
        color: "#FFEAA7"
        keyboard_shortcut: "o"
      - name: Silence
        color: "#9CA3AF"
        keyboard_shortcut: "s"

建立說話人片段

工作流

  1. 播放音訊或點選波形進行導航
  2. 在波形上點選並拖動以選擇時間範圍
  3. 按數字鍵或點選說話人標籤
  4. 該片段將被著色和標記
  5. 通過拖動邊緣調整邊界
  6. 繼續直到整個音訊都被分段

鍵盤控制

Potato 為音訊播放控制提供內建鍵盤快捷鍵，包括播放/暫停和導航。

預標註分離糾正

通常您需要糾正自動分離的結果：

yaml
data_files:
  - "data/auto_diarized.json"

資料格式：

json
{
  "id": "meeting_001",
  "audio_path": "/audio/meeting_001.wav",
  "auto_segments": [
    {"start": 0.0, "end": 3.5, "speaker": "Speaker 1"},
    {"start": 3.5, "end": 8.2, "speaker": "Speaker 2"},
    {"start": 8.2, "end": 12.0, "speaker": "Speaker 1"}
  ]
}

詳細說話人資訊

捕獲額外的說話人後設資料：

yaml
annotation_schemes:
  - annotation_type: audio_annotation
    name: speakers
    labels:
      - name: Speaker A
        color: "#FF6B6B"
      - name: Speaker B
        color: "#4ECDC4"
      - name: Speaker C
        color: "#45B7D1"
      - name: Unknown
        color: "#9CA3AF"
 
  # Speaker characteristics
  - annotation_type: radio
    name: speaker_a_gender
    description: "Speaker A Gender"
    labels:
      - Male
      - Female
      - Unknown
 
  - annotation_type: text
    name: speaker_a_role
    description: "Speaker A Role (if identifiable)"
 
  - annotation_type: radio
    name: speaker_b_gender
    description: "Speaker B Gender"
    labels:
      - Male
      - Female
      - Unknown

處理重疊語音

yaml
annotation_schemes:
  - annotation_type: audio_annotation
    name: speakers
    labels:
      - name: Speaker 1
        color: "#FF6B6B"
      - name: Speaker 2
        color: "#4ECDC4"
      - name: Overlap
        color: "#FFEAA7"

會議/訪談分離

yaml
annotation_task_name: "Meeting Diarization"
 
data_files:
  - "data/meetings.json"
 
annotation_schemes:
  # Speaker turns
  - annotation_type: audio_annotation
    name: turns
    description: "Mark each speaker turn"
    labels:
      - name: Moderator
        color: "#EF4444"
        keyboard_shortcut: "m"
      - name: Participant 1
        color: "#3B82F6"
        keyboard_shortcut: "1"
      - name: Participant 2
        color: "#10B981"
        keyboard_shortcut: "2"
      - name: Participant 3
        color: "#F59E0B"
        keyboard_shortcut: "3"
      - name: Participant 4
        color: "#8B5CF6"
        keyboard_shortcut: "4"
      - name: Unknown
        color: "#6B7280"
        keyboard_shortcut: "u"
      - name: Overlap
        color: "#FCD34D"
        keyboard_shortcut: "o"
      - name: Silence/Noise
        color: "#D1D5DB"
        keyboard_shortcut: "s"
 
  # Speech type annotation
  - annotation_type: radio
    name: speech_type
    description: "Type of speech"
    labels:
      - Statement
      - Question
      - Response
      - Interruption
      - Backchannel
 
  # Overall quality
  - annotation_type: radio
    name: recording_quality
    description: "Overall recording quality"
    labels:
      - Excellent - All speakers clear
      - Good - Most speech understandable
      - Fair - Some difficulty
      - Poor - Significant issues

輸出格式

json
{
  "id": "meeting_001",
  "audio_path": "/audio/meeting_001.wav",
  "annotations": {
    "turns": [
      {
        "start": 0.0,
        "end": 5.2,
        "label": "Moderator",
        "attributes": {
          "speech_type": "Statement"
        }
      },
      {
        "start": 5.2,
        "end": 12.8,
        "label": "Participant 1",
        "attributes": {
          "speech_type": "Response"
        }
      },
      {
        "start": 11.5,
        "end": 12.8,
        "label": "Overlap"
      }
    ],
    "recording_quality": "Good - Most speech understandable"
  }
}

分離標註技巧

  1. 先聽一遍：在標註前先熟悉說話人
  2. 記錄說話人特徵：音高、口音、說話風格
  3. 一致地處理重疊：預先確定處理策略
  4. 使用速度控制：對困難片段放慢速度
  5. 標記不確定性：需要時可以使用"未知"

下一步

完整音訊文件請見 音訊標註。