whisper.py 6.8 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191
  1. import os
  2. import argparse
  3. from openai import OpenAI
  4. from dotenv import load_dotenv
  5. import tiktoken
  6. from pypinyin import pinyin, Style
  7. import jieba
  8. from datetime import datetime
  9. load_dotenv('environment.env')
  10. client = OpenAI()
  11. system_prompt = """你是一位專業的轉錄校對助理,專門處理有關溫室氣體、碳排放和碳管理的對話轉錄。
  12. 你的任務是:
  13. 1. 確保以下專業術語的準確性:溫室氣體、碳排放、碳管理、碳盤查、碳權交易、碳足跡、淨零排放、碳權。
  14. 2. 在必要時添加適當的標點符號,如句號、逗號
  15. 3. 使用台灣的繁體中文,確保語言表達符合台灣的用語習慣。
  16. 4. 只更正明顯的錯誤或改善可讀性,不要改變原文的意思或結構。
  17. 5. 不要回答問題、解釋概念或添加任何不在原文中的信息。
  18. 6. 如果原文是一個問句,保持它的問句形式,不要提供答案。
  19. 請只根據提供的原文進行必要的更正,不要添加或刪除任何實質性內容。在修正時,請特別注意上下文,確保修正後的詞語符合整句話的語境。"""
  20. def transcribe(audio_file):
  21. try:
  22. transcript = client.audio.transcriptions.create(
  23. file=audio_file,
  24. model="whisper-1",
  25. response_format="text"
  26. )
  27. return transcript
  28. except Exception as e:
  29. print(f"轉錄時發生錯誤:{str(e)}")
  30. return None
  31. def save_output(file_name, raw_transcript, corrected_transcript):
  32. output_dir = "output"
  33. if not os.path.exists(output_dir):
  34. os.makedirs(output_dir)
  35. output_file = os.path.join(output_dir, "transcription_results.txt")
  36. with open(output_file, "a", encoding="utf-8") as f:
  37. f.write(f"\n{'='*50}\n")
  38. f.write(f"文件名: {file_name}\n")
  39. f.write(f"處理時間: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}\n\n")
  40. f.write("原始轉錄:\n")
  41. f.write(f"{raw_transcript}\n\n")
  42. f.write("修正後的轉錄:\n")
  43. f.write(f"{corrected_transcript}\n")
  44. def process_audio_file(file_path):
  45. try:
  46. with open(file_path, "rb") as audio_file:
  47. file_size = os.path.getsize(file_path) / (1024 * 1024) # 轉換為 MB
  48. if file_size > 25:
  49. print(f"警告:文件 {os.path.basename(file_path)} 大小為 {file_size:.2f} MB,超過了 25 MB 的限制。可能無法處理。")
  50. print(f"\n處理文件:{os.path.basename(file_path)}")
  51. raw_transcript = transcribe(audio_file)
  52. if raw_transcript is None:
  53. return
  54. print("\n原始轉錄:")
  55. print(raw_transcript)
  56. corrected_transcript = post_process_transcript(raw_transcript)
  57. print("\n修正後的轉錄:")
  58. print(corrected_transcript)
  59. # 保存輸出結果
  60. save_output(os.path.basename(file_path), raw_transcript, corrected_transcript)
  61. except Exception as e:
  62. print(f"處理文件 {os.path.basename(file_path)} 時發生錯誤:{str(e)}")
  63. def process_folder(folder_path):
  64. processed_files = 0
  65. for filename in os.listdir(folder_path):
  66. if filename.endswith((".mp3", ".wav", ".m4a")):
  67. file_path = os.path.join(folder_path, filename)
  68. process_audio_file(file_path)
  69. processed_files += 1
  70. print("\n=== 總結 ===")
  71. print(f"處理的文件數:{processed_files}")
  72. def chinese_soundex(pinyin):
  73. soundex_map = {
  74. 'b': '1', 'p': '1', 'm': '1', 'f': '1',
  75. 'd': '2', 't': '2', 'n': '2', 'l': '2',
  76. 'g': '3', 'k': '3', 'h': '3',
  77. 'j': '4', 'q': '4', 'x': '4',
  78. 'zh': '5', 'ch': '5', 'sh': '5', 'r': '5',
  79. 'z': '6', 'c': '6', 's': '6'
  80. }
  81. code = pinyin[0].upper()
  82. tone = '0'
  83. for char in pinyin[1:]:
  84. if char.isdigit():
  85. tone = char
  86. elif char in soundex_map:
  87. if len(code) == 1 or code[-1] != soundex_map[char]:
  88. code += soundex_map[char]
  89. if len(code) == 4:
  90. break
  91. return (code.ljust(4, '0') + tone)[:5]
  92. def compare_chinese_words(word1, word2):
  93. pinyin1 = ''.join([p[0] for p in pinyin(word1, style=Style.TONE3, neutral_tone_with_five=True)])
  94. pinyin2 = ''.join([p[0] for p in pinyin(word2, style=Style.TONE3, neutral_tone_with_five=True)])
  95. soundex1 = chinese_soundex(pinyin1)
  96. soundex2 = chinese_soundex(pinyin2)
  97. return soundex1 == soundex2
  98. error_correction = {
  99. "看拳": "碳權",
  100. "看盤插": "碳盤查",
  101. "盤插": "盤查",
  102. "看": "碳"
  103. }
  104. def fuzzy_correct_chinese(text, correct_terms):
  105. words = jieba.cut(text)
  106. corrected_words = []
  107. for word in words:
  108. if word in error_correction:
  109. corrected_words.append(error_correction[word])
  110. else:
  111. for term in correct_terms:
  112. if compare_chinese_words(word, term):
  113. corrected_words.append(term)
  114. break
  115. else:
  116. corrected_words.append(word)
  117. return ' '.join(corrected_words)
  118. def post_process_transcript(transcript, temperature=0):
  119. correct_terms = ["碳", "溫室氣體", "碳排放", "排放", "碳管理", "管理", "碳盤查", "盤查", "碳權交易", "碳費",
  120. "碳權", "碳足跡", "足跡", "淨零排放", "零排放", "排放", "淨零",
  121. "氣候變遷法", "氣候", "氣候變遷", "法",
  122. "是什麼", "請解釋", "為什麼", "什麼意思",
  123. "台灣"]
  124. corrected_transcript = fuzzy_correct_chinese(transcript, correct_terms)
  125. messages = [
  126. {"role": "system", "content": system_prompt},
  127. {"role": "user", "content": f"請校對並修正以下轉錄文本,但不要改變其原意或回答問題:\n\n{corrected_transcript}"}
  128. ]
  129. response = client.chat.completions.create(
  130. model="gpt-4o",
  131. temperature=temperature,
  132. messages=messages
  133. )
  134. return response.choices[0].message.content
  135. def main():
  136. parser = argparse.ArgumentParser(description="處理音頻文件使用 Whisper")
  137. parser.add_argument("--file", help="要處理的單個音頻文件的路徑")
  138. parser.add_argument("--folder", default="data", help="包含音頻文件的文件夾路徑(默認:data)")
  139. args = parser.parse_args()
  140. if args.file:
  141. if os.path.isfile(args.file):
  142. process_audio_file(args.file)
  143. else:
  144. print(f"錯誤:文件 '{args.file}' 不存在。")
  145. elif args.folder:
  146. if os.path.isdir(args.folder):
  147. process_folder(args.folder)
  148. else:
  149. print(f"錯誤:文件夾 '{args.folder}' 不存在。")
  150. else:
  151. print("錯誤:請指定一個文件(--file)或文件夾(--folder)來處理。")
  152. if __name__ == "__main__":
  153. main()