pip install openai-whisper transformers pydub librosa tqdm torch ffmpeg-python modelscope

from modelscope import snapshot_download

# 下載模型到當(dāng)前目錄
model_dir = snapshot_download('Helsinki-NLP/opus-mt-en-zh', cache_dir='./')

print(f"? 模型已下載到當(dāng)前目錄: {model_dir}")

model_dir = snapshot_download('Helsinki-NLP/opus-mt-zh-en', cache_dir='./')

print(f"? 模型已下載到當(dāng)前目錄: {model_dir}")

import whisper
import warnings
from datetime import timedelta
from tqdm import tqdm
import librosa
import time
from transformers import pipeline
import torch
from pydub import AudioSegment
import os
import ffmpeg

# 忽略警告信息
warnings.filterwarnings("ignore", category=UserWarning)

# 輸入視頻路徑（可修改）
input_video = r"./Gesture Drawing Practice _ 20 and 40 sec. poses.mp4"
# 輸出音頻路徑
output_audio = "Gesture Drawing Practice _ 20 and 40 sec. poses.wav"

# =========================
# Step 1: 提取視頻中的音頻
# =========================
ffmpeg.input(input_video).output(output_audio, ac=1, ar=16000).run()

# =========================
# Step 2: 時(shí)間戳格式化函數(shù)
# =========================
def format_timestamp(seconds):
    """將秒數(shù)轉(zhuǎn)換為 SRT 時(shí)間格式（HH:MM:SS,mmm）"""
    td = timedelta(seconds=seconds)
    hours, remainder = divmod(td.seconds, 3600)
    minutes, seconds = divmod(remainder, 60)
    milliseconds = int(td.microseconds / 1000)
    return f"{hours:02d}:{minutes:02d}:{seconds:02d},{milliseconds:03d}"

# =========================
# Step 3: 驗(yàn)證音頻文件是否有效
# =========================
def validate_audio(audio_file):
    """檢查音頻文件是否有效，并返回其基本信息"""
    try:
        audio = AudioSegment.from_file(audio_file)
        duration = len(audio) / 1000.0  # 秒
        sample_rate = audio.frame_rate
        channels = audio.channels
        print(f"? 音頻驗(yàn)證成功：時(shí)長(zhǎng)={duration:.2f}s, 采樣率={sample_rate}Hz, 聲道={channels}")
        return True, duration
    except Exception as e:
        print(f"? 音頻驗(yàn)證失敗: {e}")
        return False, 0

# =========================
# Step 4: 初始化翻譯模型
# =========================
print("?? 正在加載翻譯模型，請(qǐng)稍候...")
translator_en_to_zh = pipeline("translation", model="./Helsinki-NLP/opus-mt-en-zh")  # 英譯中
translator_zh_to_en = pipeline("translation", model="./Helsinki-NLP/opus-mt-zh-en")  # 中譯英
print("? 翻譯模型加載完成。")

# =========================
# Step 5: 驗(yàn)證音頻文件
# =========================
audio_file = output_audio
is_valid, duration = validate_audio(audio_file)
if not is_valid:
    raise ValueError(f"無(wú)效的音頻文件: {audio_file}")

# 備用方案：使用 librosa 檢查時(shí)長(zhǎng)
try:
    duration = librosa.get_duration(path=audio_file)
except Exception as e:
    print(f"Librosa 檢查失敗: {e}")

# =========================
# Step 6: 初始化 Whisper 模型
# =========================
print("??? 正在加載 Whisper 模型...")
try:
    model = whisper.load_model("base")  # 可選 tiny, base, small, medium, large
except Exception as e:
    print(f"加載模型失敗: {e}，回退到 'medium' 模型。")
    model = whisper.load_model("medium")

# 進(jìn)度條初始化
progress_bar = tqdm(total=100, desc="語(yǔ)音識(shí)別中", unit="%")

# =========================
# Step 7: 執(zhí)行語(yǔ)音識(shí)別（帶進(jìn)度）
# =========================
def transcribe_with_progress(model, audio_file, language=None):
    start_time = time.time()
    
    # 自動(dòng)檢測(cè)語(yǔ)言
    if language is None:
        try:
            audio = whisper.load_audio(audio_file)
            mel = whisper.log_mel_spectrogram(audio, n_mels=80).to(model.device)
            _, probs = model.detect_language(mel)
            detected_language = max(probs, key=probs.get)
            print(f"?? 自動(dòng)檢測(cè)語(yǔ)言: {detected_language}")
        except Exception as e:
            print(f"語(yǔ)言檢測(cè)失敗: {e}，默認(rèn)使用英語(yǔ)(en)。")
            detected_language = "en"
    else:
        detected_language = language
        print(f"?? 使用指定語(yǔ)言: {detected_language}")
    
    # 執(zhí)行轉(zhuǎn)錄
    try:
        result = model.transcribe(audio_file, language=detected_language)
    except Exception as e:
        print(f"轉(zhuǎn)錄失敗: {e}")
        raise
    
    # 更新進(jìn)度條
    progress_bar.update(100 - progress_bar.n)
    progress_bar.close()
    
    elapsed_time = time.time() - start_time
    print(f"? 轉(zhuǎn)錄完成，用時(shí) {elapsed_time:.2f} 秒。")
    
    return result, detected_language

# 執(zhí)行轉(zhuǎn)錄（可指定語(yǔ)言）
result, detected_language = transcribe_with_progress(model, audio_file, language="en")

# =========================
# Step 8: 生成中英雙語(yǔ)字幕文件
# =========================
srt_path = f"{output_audio}.srt"
with open(srt_path, "w", encoding="utf-8") as f:
    for i, segment in enumerate(result["segments"]):
        start_time = format_timestamp(segment["start"])
        end_time = format_timestamp(segment["end"])
        text = segment["text"].strip()
        
        # 判斷語(yǔ)言并翻譯
        if detected_language == "zh":
            # 中文音頻：原文中文 + 翻譯英文
            zh_text = text
            en_text = translator_zh_to_en(text)[0]["translation_text"]
        else:
            # 英語(yǔ)音頻：原文英文 + 翻譯中文
            zh_text = translator_en_to_zh(text)[0]["translation_text"]
            en_text = text
        
        # 寫入 SRT 文件（中文在上，英文在下）
        f.write(f"{i+1}\n")
        f.write(f"{start_time} --> {end_time}\n")
        f.write(f"{zh_text}\n{en_text}\n\n")

print(f"?? 字幕文件生成成功：{srt_path}")