ASR/whisper.py

import os
import argparse
from openai import OpenAI
from dotenv import load_dotenv
import jieba
from datetime import datetime
from pypinyin import pinyin, Style

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 save_output(file_name, raw_transcript, corrected_transcript):
    output_dir = "output"
    os.makedirs(output_dir, exist_ok=True)
    
    output_file = os.path.join(output_dir, "transcription_results.txt")
    
    with open(output_file, "a", encoding="utf-8") as f:
        f.write(f"\n{'='*50}\n")
        f.write(f"文件名: {file_name}\n")
        f.write(f"處理時間: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}\n\n")
        f.write("原始轉錄:\n")
        f.write(f"{raw_transcript}\n\n")
        f.write("修正後的轉錄:\n")
        f.write(f"{corrected_transcript}\n")

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)

            save_output(os.path.basename(file_path), raw_transcript, 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.lower().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_str):
    soundex_map = {
        'b': '1', 'p': '1', 'm': '1', 'f': '1',
        'd': '2', 'n': '2', 'l': '2',
        'g': '3', 'k': '3', 'h': '3', 't': '3',
        'j': '4', 'q': '4', 'x': '4',
        'zh': '5', 'ch': '5', 'sh': '5', 'r': '5',
        'z': '6', 'c': '6', 's': '6',
        'an': '7', 'ang': '7', 'en': '8', 'eng': '8', 'in': '8', 'ing': '8',
        'ong': '9', 'un': '9', 'uan': '9',
        'i': 'A', 'u': 'A', 'v': 'A',  # 'v' is used for 'ü' in some systems
        'e': 'B', 'o': 'B',
    }
    
    code = ''
    tone = '0'
    i = 0
    while i < len(pinyin_str):
        if pinyin_str[i:i+2] in soundex_map:
            code += soundex_map[pinyin_str[i:i+2]]
            i += 2
        elif pinyin_str[i] in soundex_map:
            code += soundex_map[pinyin_str[i]]
            i += 1
        elif pinyin_str[i].isdigit():
            tone = pinyin_str[i]
            i += 1
        else:
            i += 1
    
    code = code[:1] + ''.join(sorted(set(code[1:])))
    return (code[:3] + tone).ljust(4, '0')

def compare_chinese_words(word1, word2, tone_sensitive=True):
    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)
    
    if tone_sensitive:
        return soundex1 == soundex2
    else:
        return soundex1[:3] == soundex2[:3]

error_correction = {
    "看拳": "碳權",
    "看盤插": "碳盤查",
    "盤插": "盤查",
    "看": "碳",
    "看權": "碳權"
}

def set_jieba_dictionary(dict_path):
    if os.path.exists(dict_path):
        jieba.set_dictionary(dict_path)
        jieba.initialize()
        print(f"Set jieba dictionary to {dict_path}")
    else:
        print(f"Dictionary file {dict_path} not found. Using default dictionary.")

def fuzzy_correct_chinese(text, correct_terms):
    words = jieba.lcut(text)
    corrected_words = []
    for word in words:
        word_pinyin = ''.join([p[0] for p in pinyin(word, style=Style.NORMAL)])
        print(f"Term: {word}, Pinyin: {word_pinyin}")  # Added print statement
        if word in error_correction:
            corrected_words.append(error_correction[word])
        else:
            for term in correct_terms:
                if compare_chinese_words(word, term, tone_sensitive=True):
                    print(f"corrected: {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 = ["溫室氣體", "碳排放", "碳管理", "碳盤查", "碳權交易", "碳費",
                     "碳權", "碳足跡", "足跡", "淨零排放", "零排放", "淨零",
                     "氣候變遷法", "氣候", "氣候變遷",
                     "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-4",
        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）")
    parser.add_argument("--dict", default="dictionary/dictionary_reviewed_rows_weighted.txt")
    args = parser.parse_args()

    set_jieba_dictionary(args.dict)

    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()