123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166 |
- import os
- import argparse
- from openai import OpenAI
- from dotenv import load_dotenv
- import tiktoken
- from pypinyin import pinyin, Style
- import jieba
- load_dotenv('environment.env')
- client = OpenAI()
- 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。
- 你的任務是:
- 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。
- 2. 在必要時添加適當的標點符號,如句號、逗號
- 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。
- 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。
- 5. 不要回答問題、解釋概念或添加任何不在原文中的信息。
- 6. 如果原文是一個問句,保持它的問句形式,不要提供答案。
- 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。"""
- def transcribe(audio_file):
- try:
- transcript = client.audio.transcriptions.create(
- file=audio_file,
- model="whisper-1",
- response_format="text"
- )
- return transcript
- except Exception as e:
- print(f"轉錄時發生錯誤:{str(e)}")
- return None
- def process_audio_file(file_path):
- try:
- with open(file_path, "rb") as audio_file:
- file_size = os.path.getsize(file_path) / (1024 * 1024) # 轉換為 MB
- if file_size > 25:
- 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。")
- print(f"\n處理文件:{os.path.basename(file_path)}")
- raw_transcript = transcribe(audio_file)
- if raw_transcript is None:
- return
- print("\n原始轉錄:")
- print(raw_transcript)
- corrected_transcript = post_process_transcript(raw_transcript)
- print("\n修正後的轉錄:")
- print(corrected_transcript)
- except Exception as e:
- print(f"處理文件 {os.path.basename(file_path)} 時發生錯誤:{str(e)}")
- def process_folder(folder_path):
- processed_files = 0
- for filename in os.listdir(folder_path):
- if filename.endswith((".mp3", ".wav", ".m4a")):
- file_path = os.path.join(folder_path, filename)
- process_audio_file(file_path)
- processed_files += 1
- print("\n=== 總結 ===")
- print(f"處理的文件數:{processed_files}")
- def chinese_soundex(pinyin):
- soundex_map = {
- 'b': '1', 'p': '1', 'm': '1', 'f': '1',
- 'd': '2', 't': '2', 'n': '2', 'l': '2',
- 'g': '3', 'k': '3', 'h': '3',
- 'j': '4', 'q': '4', 'x': '4',
- 'zh': '5', 'ch': '5', 'sh': '5', 'r': '5',
- 'z': '6', 'c': '6', 's': '6'
- }
-
- code = pinyin[0].upper()
- tone = '0'
-
- for char in pinyin[1:]:
- if char.isdigit():
- tone = char
- elif char in soundex_map:
- if len(code) == 1 or code[-1] != soundex_map[char]:
- code += soundex_map[char]
- if len(code) == 4:
- break
-
- return (code.ljust(4, '0') + tone)[:5]
- def compare_chinese_words(word1, word2):
- pinyin1 = ''.join([p[0] for p in pinyin(word1, style=Style.TONE3, neutral_tone_with_five=True)])
- pinyin2 = ''.join([p[0] for p in pinyin(word2, style=Style.TONE3, neutral_tone_with_five=True)])
-
- soundex1 = chinese_soundex(pinyin1)
- soundex2 = chinese_soundex(pinyin2)
-
- return soundex1 == soundex2
- error_correction = {
- "看拳": "碳權",
- "看盤插": "碳盤查",
- "盤插": "盤查",
- "看": "碳"
- }
- def fuzzy_correct_chinese(text, correct_terms):
- words = jieba.cut(text)
- corrected_words = []
- for word in words:
- if word in error_correction:
- corrected_words.append(error_correction[word])
- else:
- for term in correct_terms:
- if compare_chinese_words(word, term):
- corrected_words.append(term)
- break
- else:
- corrected_words.append(word)
- return ' '.join(corrected_words)
- def post_process_transcript(transcript, temperature=0):
- 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",
- "碳權", "碳足跡", "足跡", "淨零排放", "零排放", "排放", "淨零",
- "氣候變遷法", "氣候", "氣候變遷", "法",
- "是什麼", "請解釋", "為什麼", "什麼意思",
- "台灣"]
-
- corrected_transcript = fuzzy_correct_chinese(transcript, correct_terms)
-
- messages = [
- {"role": "system", "content": system_prompt},
- {"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}"}
- ]
- response = client.chat.completions.create(
- model="gpt-4o",
- temperature=temperature,
- messages=messages
- )
- return response.choices[0].message.content
- def main():
- parser = argparse.ArgumentParser(description="處理音頻文件使用 Whisper")
- parser.add_argument("--file", help="要處理的單個音頻文件的路徑")
- parser.add_argument("--folder", default="data", help="包含音頻文件的文件夾路徑(默認:data)")
- args = parser.parse_args()
- if args.file:
- if os.path.isfile(args.file):
- process_audio_file(args.file)
- else:
- print(f"錯誤:文件 '{args.file}' 不存在。")
- elif args.folder:
- if os.path.isdir(args.folder):
- process_folder(args.folder)
- else:
- print(f"錯誤:文件夾 '{args.folder}' 不存在。")
- else:
- print("錯誤:請指定一個文件(--file)或文件夾(--folder)來處理。")
- if __name__ == "__main__":
- main()
|