Transcription Automation

文字起こしの自動化

音声・動画の文字起こしとコンテンツ処理を自動化する包括的なスキルです。

基本的なワークフロー

1. 文字起こしパイプライン

TRANSCRIPTION FLOW:
┌─────────────────┐
│  Audio/Video    │
│     Input       │
└────────┬────────┘
         ▼
┌─────────────────┐
│  Pre-Processing │
│  - Convert      │
│  - Enhance      │
│  - Split        │
└────────┬────────┘
         ▼
┌─────────────────┐
│  Transcription  │
│  - STT Engine   │
│  - Diarization  │
└────────┬────────┘
         ▼
┌─────────────────┐
│ Post-Processing │
│  - Format       │
│  - Timestamps   │
│  - Speakers     │
└────────┬────────┘
         ▼
┌─────────────────┐
│     Output      │
│  - Text/SRT/VTT │
│  - Summary      │
└─────────────────┘

2. 文字起こしの設定

transcription_config:
  engine: whisper  # whisper, assembly_ai, deepgram
  
  audio_settings:
    sample_rate: 16000
    channels: mono
    format: wav
    
  transcription:
    language: auto  # or specific: en, zh, es
    model: large  # tiny, base, small, medium, large
    task: transcribe  # transcribe or translate
    
  features:
    speaker_diarization: true
    word_timestamps: true
    punctuation: true
    profanity_filter: false
    
  output:
    formats:
      - txt
      - srt
      - vtt
      - json
    include_confidence: true
    include_timestamps: true

会議の文字起こし

議事録テンプレート

meeting_transcript:
  metadata:
    title: "{{meeting_title}}"
    date: "{{date}}"
    duration: "{{duration}}"
    attendees: "{{speakers}}"
    
  output_template: |
    # {{title}}
    
    **Date:** {{date}}
    **Duration:** {{duration}}
    **Attendees:** {{attendees}}
    
    ## Summary
    {{ai_summary}}
    
    ## Key Points
    {{#each key_points}}
    - {{this}}
    {{/each}}
    
    ## Action Items
    {{#each action_items}}
    - [ ] {{task}} - @{{assignee}} - Due: {{due_date}}
    {{/each}}
    
    ## Full Transcript
    {{#each segments}}
    **[{{timestamp}}] {{speaker}}:** {{text}}
    
    {{/each}}

話者ダイアライゼーション

diarization_config:
  min_speakers: 2
  max_speakers: 10
  
  speaker_labels:
    - name: "Speaker 1"
      voice_sample: "sample_1.wav"  # Optional
    - name: "Speaker 2"
      voice_sample: "sample_2.wav"
      
  output_format:
    speaker_prefix: true
    speaker_timestamps: true
    
  example_output: |
    [00:00:05] SPEAKER_1: Welcome everyone to today's meeting.
    [00:00:12] SPEAKER_2: Thanks for having us.
    [00:00:18] SPEAKER_1: Let's start with the agenda.

字幕の生成

SRT 形式

subtitle_config:
  format: srt
  
  timing:
    max_duration: 7  # seconds per subtitle
    min_gap: 0.1     # seconds between subtitles
    chars_per_line: 42
    max_lines: 2
    
  style:
    case: sentence  # sentence, upper, lower
    numbers: words  # words, digits
    
  example_output: |
    1
    00:00:05,000 --> 00:00:08,500
    Welcome to today's presentation
    about transcription automation.
    
    2
    00:00:09,000 --> 00:00:12,000
    Let me start by explaining
    the basic concepts.

VTT 形式

vtt_config:
  format: vtt
  
  features:
    cue_settings: true
    styling: true
    
  example_output: |
    WEBVTT
    
    00:00:05.000 --> 00:00:08.500 align:center
    Welcome to today's presentation
    about transcription automation.
    
    00:00:09.000 --> 00:00:12.000 align:center
    <v Speaker 1>Let me start by explaining
    the basic concepts.

統合ワークフロー

Zoom との統合

zoom_transcription:
  trigger:
    event: recording_completed
    
  workflow:
    - step: download_recording
      source: zoom_cloud
      
    - step: transcribe
      engine: whisper
      language: auto
      
    - step: diarize
      identify_speakers: true
      
    - step: generate_notes
      template: meeting_notes
      include_summary: true
      extract_action_items: true
      
    - step: distribute
      destinations:
        - notion_page
        - slack_channel
        - email_attendees

YouTube との統合

youtube_subtitles:
  trigger:
    event: video_uploaded
    
  workflow:
    - step: download_audio
      source: youtube_video
      
    - step: transcribe
      engine: whisper
      task: transcribe
      
    - step: generate_subtitles
      formats: [srt, vtt]
      
    - step: translate
      target_languages: [es, zh, ja, de, fr]
      
    - step: upload_subtitles
      destination: youtube
      as_cc: true

ポッドキャストの処理

podcast_workflow:
  input:
    source: rss_feed
    format: audio/mp3
    
  processing:
    - transcribe:
        engine: whisper
        model: large
        
    - generate_chapters:
        detect_topics: true
        min_duration: 60  # seconds
        
    - create_show_notes:
        summarize: true
        extract_links: true
        highlight_quotes: true
        
    - create_searchable_index:
        full_text: true
        timestamps: true
        
  output:
    - transcript_txt
    - chapters_json
    - show_notes_md
    - search_index

言語サポート

多言語の文字起こし

multilingual:
  auto_detect: true
  
  supported_languages:
    - code: en
      name: English
      model: large
      
    - code: zh
      name: Chinese
      model: large
      
    - code: es
      name: Spanish
      model: large
      
    - code: ja
      name: Japanese
      model: medium
      
  translation:
    enabled: true
    target: en
    preserve_original: true

コードスイッチング

code_switching:
  enabled: true
  primary_language: en
  secondary_languages: [zh, es]
  
  output: |
    [00:01:23] The next topic is about 人工智能,
    which has been muy importante in recent years.
    
  handling:
    detect_language_per_segment: true
    tag_language_switches: true

品質の向上

後処理

post_processing:
  text_cleanup:
    - remove_filler_words: ["um", "uh", "like"]
    - fix_common_errors: true
    - normalize_numbers: true
    
  formatting:
    - add_punctuation: true
    - capitalize_sentences: true
    - paragraph_breaks: true
    
  speaker_attribution:
    - merge_short_segments: true
    - min_segment_duration: 1.0
    
  output_enhancement:
    - add_timestamps: true
    - highlight_keywords: true
    - generate_summary: true

精度指標

TRANSCRIPTION QUALITY REPORT
═══════════════════════════════════════

File: meeting_2024_01_15.mp3
Duration: 45:32
Engine: Whisper Large

METRICS:
Word Error Rate (WER):  4.2%
Character Error Rate:   2.8%
Confidence Score:       0.94

SPEAKER DIARIZATION:
Speakers Detected: 4
Diarization Accuracy: 91%

PROCESSING TIME:
Total: 8m 23s
Real-time Factor: 0.18x

DETECTED ISSUES:
• Low confidence at 12:34 (background noise)
• Overlapping speech at 23:45
• Unknown speaker at 34:12

API の例

OpenAI Whisper

import openai

# Transcribe audio
with open("meeting.mp3", "rb") as audio_file:
    transcript = openai.Audio.transcribe(
        model="whisper-1",
        file=audio_file,
        response_format="verbose_json",
        timestamp_granularities=["word", "segment"]
    )

# Access results
for segment in transcript.segments:
    print(f"[{segment.start:.2f}] {segment.text}")

AssemblyAI

import assemblyai as aai

transcriber = aai.Transcriber()

config = aai.TranscriptionConfig(
    speaker_labels=True,
    auto_chapters=True,
    entity_detection=True
)

transcript = transcriber.transcribe(
    "https://example.com/meeting.mp3",
    config=config
)

for utterance in transcript.utterances:
    print(f"Speaker {utterance.speaker}: {utterance.text}")

ベストプラクティス

1. 高品質な音声：入力がクリアなら、出力も向上します
2. 適切なモデルを選択：速度と精度のバランスを取ります
3. ダイアライゼーションを使用：話者を明確に識別します
4. 後処理を実施：自動生成された出力を整えます
5. 重要な内容を確認：人によるレビューが重要です
6. プライバシーを考慮：機密性の高いコンテンツを適切に扱います
7. 効率的に保存：圧縮してインデックス化します
8. コンテキストを提供：語彙のヒントが役立ちます