whisper.py 6.0 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166
  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. load_dotenv('environment.env')
  9. client = OpenAI()
  10. 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。
  11. 你的任務是:
  12. 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。
  13. 2. 在必要時添加適當的標點符號,如句號、逗號
  14. 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。
  15. 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。
  16. 5. 不要回答問題、解釋概念或添加任何不在原文中的信息。
  17. 6. 如果原文是一個問句,保持它的問句形式,不要提供答案。
  18. 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。"""
  19. def transcribe(audio_file):
  20. try:
  21. transcript = client.audio.transcriptions.create(
  22. file=audio_file,
  23. model="whisper-1",
  24. response_format="text"
  25. )
  26. return transcript
  27. except Exception as e:
  28. print(f"轉錄時發生錯誤:{str(e)}")
  29. return None
  30. def process_audio_file(file_path):
  31. try:
  32. with open(file_path, "rb") as audio_file:
  33. file_size = os.path.getsize(file_path) / (1024 * 1024) # 轉換為 MB
  34. if file_size > 25:
  35. 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。")
  36. print(f"\n處理文件:{os.path.basename(file_path)}")
  37. raw_transcript = transcribe(audio_file)
  38. if raw_transcript is None:
  39. return
  40. print("\n原始轉錄:")
  41. print(raw_transcript)
  42. corrected_transcript = post_process_transcript(raw_transcript)
  43. print("\n修正後的轉錄:")
  44. print(corrected_transcript)
  45. except Exception as e:
  46. print(f"處理文件 {os.path.basename(file_path)} 時發生錯誤:{str(e)}")
  47. def process_folder(folder_path):
  48. processed_files = 0
  49. for filename in os.listdir(folder_path):
  50. if filename.endswith((".mp3", ".wav", ".m4a")):
  51. file_path = os.path.join(folder_path, filename)
  52. process_audio_file(file_path)
  53. processed_files += 1
  54. print("\n=== 總結 ===")
  55. print(f"處理的文件數:{processed_files}")
  56. def chinese_soundex(pinyin):
  57. soundex_map = {
  58. 'b': '1', 'p': '1', 'm': '1', 'f': '1',
  59. 'd': '2', 't': '2', 'n': '2', 'l': '2',
  60. 'g': '3', 'k': '3', 'h': '3',
  61. 'j': '4', 'q': '4', 'x': '4',
  62. 'zh': '5', 'ch': '5', 'sh': '5', 'r': '5',
  63. 'z': '6', 'c': '6', 's': '6'
  64. }
  65. code = pinyin[0].upper()
  66. tone = '0'
  67. for char in pinyin[1:]:
  68. if char.isdigit():
  69. tone = char
  70. elif char in soundex_map:
  71. if len(code) == 1 or code[-1] != soundex_map[char]:
  72. code += soundex_map[char]
  73. if len(code) == 4:
  74. break
  75. return (code.ljust(4, '0') + tone)[:5]
  76. def compare_chinese_words(word1, word2):
  77. pinyin1 = ''.join([p[0] for p in pinyin(word1, style=Style.TONE3, neutral_tone_with_five=True)])
  78. pinyin2 = ''.join([p[0] for p in pinyin(word2, style=Style.TONE3, neutral_tone_with_five=True)])
  79. soundex1 = chinese_soundex(pinyin1)
  80. soundex2 = chinese_soundex(pinyin2)
  81. return soundex1 == soundex2
  82. error_correction = {
  83. "看拳": "碳權",
  84. "看盤插": "碳盤查",
  85. "盤插": "盤查",
  86. "看": "碳"
  87. }
  88. def fuzzy_correct_chinese(text, correct_terms):
  89. words = jieba.cut(text)
  90. corrected_words = []
  91. for word in words:
  92. if word in error_correction:
  93. corrected_words.append(error_correction[word])
  94. else:
  95. for term in correct_terms:
  96. if compare_chinese_words(word, term):
  97. corrected_words.append(term)
  98. break
  99. else:
  100. corrected_words.append(word)
  101. return ' '.join(corrected_words)
  102. def post_process_transcript(transcript, temperature=0):
  103. 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",
  104. "碳權", "碳足跡", "足跡", "淨零排放", "零排放", "排放", "淨零",
  105. "氣候變遷法", "氣候", "氣候變遷", "法",
  106. "是什麼", "請解釋", "為什麼", "什麼意思",
  107. "台灣"]
  108. corrected_transcript = fuzzy_correct_chinese(transcript, correct_terms)
  109. messages = [
  110. {"role": "system", "content": system_prompt},
  111. {"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}"}
  112. ]
  113. response = client.chat.completions.create(
  114. model="gpt-4o",
  115. temperature=temperature,
  116. messages=messages
  117. )
  118. return response.choices[0].message.content
  119. def main():
  120. parser = argparse.ArgumentParser(description="處理音頻文件使用 Whisper")
  121. parser.add_argument("--file", help="要處理的單個音頻文件的路徑")
  122. parser.add_argument("--folder", default="data", help="包含音頻文件的文件夾路徑(默認:data)")
  123. args = parser.parse_args()
  124. if args.file:
  125. if os.path.isfile(args.file):
  126. process_audio_file(args.file)
  127. else:
  128. print(f"錯誤:文件 '{args.file}' 不存在。")
  129. elif args.folder:
  130. if os.path.isdir(args.folder):
  131. process_folder(args.folder)
  132. else:
  133. print(f"錯誤:文件夾 '{args.folder}' 不存在。")
  134. else:
  135. print("錯誤:請指定一個文件(--file)或文件夾(--folder)來處理。")
  136. if __name__ == "__main__":
  137. main()