sharepoint-transcribe/simplify_transcript.py

#!/usr/bin/env python3
"""
简化 Amazon Transcribe JSON 输出，合并为按说话人分段的文本。
适合直接喂给 LLM 做会议总结。

用法: python simplify_transcript.py testjob.json [-o output.txt]
"""

import json
import sys
import argparse


def load_transcribe(path: str) -> dict:
    with open(path, 'r', encoding='utf-8') as f:
        return json.load(f)


def simplify(data: dict) -> str:
    """将 Amazon Transcribe JSON 合并为 Speaker + 文本段落"""
    items = data['results']['items']
    
    # 构建段落: 按说话人变化分段
    paragraphs = []
    current_speaker = None
    current_text = []
    current_start = None
    
    for item in items:
        speaker = item.get('speaker_label', '')
        content = item['alternatives'][0]['content']
        item_type = item['type']  # pronunciation / punctuation
        
        # 标点符号不换说话人，直接追加
        if item_type == 'punctuation':
            if current_text:
                current_text.append(content)
            continue
        
        # 说话人变了，保存当前段落
        if speaker != current_speaker and current_text:
            paragraphs.append({
                'speaker': current_speaker,
                'start': current_start,
                'text': ''.join(current_text).strip(),
            })
            current_text = []
        
        if speaker != current_speaker:
            current_speaker = speaker
            current_start = item.get('start_time', '')
        
        # 中文不加空格，英文/数字前加空格
        if current_text and not content.startswith((',', '。', '？', '！', '、', '：', '；')):
            # 判断是否需要空格 (前一个和当前都是 ASCII 才加)
            prev = current_text[-1] if current_text else ''
            if prev and prev[-1].isascii() and content[0].isascii():
                current_text.append(' ')
        current_text.append(content)
    
    # 最后一段
    if current_text:
        paragraphs.append({
            'speaker': current_speaker,
            'start': current_start,
            'text': ''.join(current_text).strip(),
        })
    
    # 格式化输出
    lines = []
    speaker_count = len(set(p['speaker'] for p in paragraphs))
    lines.append(f"# 会议转录 ({speaker_count} 位发言人, {len(paragraphs)} 段)")
    lines.append("")
    
    for p in paragraphs:
        start = p['start']
        if start:
            secs = float(start)
            mins = int(secs // 60)
            s = int(secs % 60)
            ts = f"[{mins:02d}:{s:02d}]"
        else:
            ts = ""
        
        lines.append(f"**{p['speaker']}** {ts}")
        lines.append(p['text'])
        lines.append("")
    
    return '\n'.join(lines)


def main():
    parser = argparse.ArgumentParser(description="简化 Amazon Transcribe JSON")
    parser.add_argument("input", help="Amazon Transcribe JSON 文件")
    parser.add_argument("-o", "--output", help="输出文件 (默认: 同名.txt)")
    args = parser.parse_args()
    
    data = load_transcribe(args.input)
    result = simplify(data)
    
    out_path = args.output or args.input.rsplit('.', 1)[0] + '_simplified.txt'
    with open(out_path, 'w', encoding='utf-8') as f:
        f.write(result)
    
    print(f"已保存: {out_path}")
    print(f"大小: {len(result)} 字符 (原始 JSON: {len(open(args.input, encoding='utf-8').read())} 字符)")
    print(f"压缩比: {len(result) / len(open(args.input, encoding='utf-8').read()) * 100:.1f}%")


if __name__ == "__main__":
    main()