app.py

import torch
import os
import gradio as gr
from pyannote.audio import Pipeline
from pydub import AudioSegment
from spaces import GPU
from io import BytesIO

# 获取 Hugging Face 认证令牌
HF_TOKEN = os.environ.get("HUGGINGFACE_READ_TOKEN")
pipeline = None

# 尝试加载 pyannote 模型
try:
    pipeline = Pipeline.from_pretrained(
        "pyannote/speaker-diarization-3.1", use_auth_token=HF_TOKEN
    )
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    pipeline.to(device)
except Exception as e:
    print(f"Error initializing pipeline: {e}")
    pipeline = None

# 时间戳转换为秒
def timestamp_to_seconds(timestamp):
    h, m, s = map(float, timestamp.split(':'))
    return 3600 * h + 60 * m + s

def convert_to_wav(audio_file):
    try:
        # 使用 pydub 打开上传的音频文件
        audio = AudioSegment.from_file(audio_file)
        
        # 创建一个 BytesIO 对象以存储转换后的音频
        wav_output = BytesIO()
        
        # 将音频转换为 wav 格式并存储在 BytesIO 对象中
        audio.export(wav_output, format="wav")
        
        # 将 BytesIO 对象的位置重置为开始位置，以便之后可以读取
        wav_output.seek(0)
        
        return wav_output  # 返回转换后的 wav 音频
    except Exception as e:
        return f"音频转换失败: {e}"

# 音频拼接函数：拼接目标音频和混合音频，返回目标音频的起始时间和结束时间作为字典
def combine_audio_with_time(target_audio, mixed_audio):
    if pipeline is None:
        return "错误: 模型未初始化"
    
    # 打印文件路径，确保文件正确传递
    print(f"目标音频文件路径: {target_audio}")
    print(f"混合音频文件路径: {mixed_audio}")

    target_audio = convert_to_wav(target_audio)
    mixed_audio = convert_to_wav(mixed_audio)

    print(f"目标音频文件路径: {target_audio}")
    print(f"混合音频文件路径: {mixed_audio}")
    
    # 加载目标说话人的样本音频
    try:
        target_audio_segment = AudioSegment.from_wav(target_audio)
    except Exception as e:
        return f"加载目标音频时出错: {e}"
    
    # 加载混合音频
    try:
        mixed_audio_segment = AudioSegment.from_wav(mixed_audio)
    except Exception as e:
        return f"加载混合音频时出错: {e}"

    # 记录目标说话人音频的时间点（精确到0.01秒）
    target_start_time = len(mixed_audio_segment) / 1000  # 秒为单位，精确到 0.01 秒

    # 目标音频的结束时间（拼接后的音频长度）
    target_end_time = target_start_time + len(target_audio_segment) / 1000  # 秒为单位

    # 将目标说话人的音频片段添加到混合音频的最后
    final_audio = mixed_audio_segment + target_audio_segment
    final_audio.export("final_output.wav", format="wav")
    
    # 返回目标音频的起始时间和结束时间
    return {"start_time": target_start_time, "end_time": target_end_time}

# 使用 pyannote/speaker-diarization 对拼接后的音频进行说话人分离
@GPU(duration=60 * 2)  # 使用 GPU 加速，限制执行时间为 120 秒
def diarize_audio(temp_file):
    if pipeline is None:
        return "错误: 模型未初始化"
    
    try:
        diarization = pipeline(temp_file)
        print("说话人分离结果:")
        for turn, _, speaker in diarization.itertracks(yield_label=True):
            print(f"[{turn.start:.3f} --> {turn.end:.3f}] {speaker}")
        return diarization
    except Exception as e:
        return f"处理音频时出错: {e}"

# 查找最匹配的说话人
def find_best_matching_speaker(target_start_time, target_end_time, diarization):
    best_match = None
    max_overlap = 0
    
    # 遍历所有说话人时间段，计算与目标音频的重叠部分
    for turn, _, speaker in diarization.itertracks(yield_label=True):
        start = turn.start
        end = turn.end
        
        # 计算重叠部分的开始和结束时间
        overlap_start = max(start, target_start_time)
        overlap_end = min(end, target_end_time)
        
        # 如果有重叠部分，计算重叠的持续时间
        if overlap_end > overlap_start:
            overlap_duration = overlap_end - overlap_start
            
            # 如果当前重叠部分更大，则更新最匹配的说话人
            if overlap_duration > max_overlap:
                max_overlap = overlap_duration
                best_match = speaker
    
    return best_match, max_overlap

# 获取目标说话人的时间段（排除目标音频时间段）
def get_speaker_segments(diarization, target_start_time, target_end_time, final_audio_length):
    speaker_segments = {}
    
    # 遍历所有说话人时间段
    for turn, _, speaker in diarization.itertracks(yield_label=True):
        start = turn.start
        end = turn.end
        
        # 如果时间段与目标音频有重叠，需要截断
        if start < target_end_time and end > target_start_time:
            # 记录被截断的时间段
            if start < target_start_time:
                # 目标音频开始前的时间段
                speaker_segments.setdefault(speaker, []).append((start, min(target_start_time, end)))
            
            if end > target_end_time:
                # 目标音频结束后的时间段
                speaker_segments.setdefault(speaker, []).append((max(target_end_time, start), min(end, final_audio_length)))
        else:
            # 完全不与目标音频重叠的时间段
            if end <= target_start_time or start >= target_end_time:
                speaker_segments.setdefault(speaker, []).append((start, end))
    
    return speaker_segments

# 处理音频文件并返回输出
def process_audio(target_audio, mixed_audio):
    print(f"处理音频：目标音频: {target_audio}, 混合音频: {mixed_audio}")
    
    # 进行音频拼接并返回目标音频的起始和结束时间（作为字典）
    time_dict = combine_audio_with_time(target_audio, mixed_audio)
    
    # 如果音频拼接出错，返回错误信息
    if isinstance(time_dict, str):
        return time_dict
    
    # 执行说话人分离
    diarization_result = diarize_audio("final_output.wav")
    
    if isinstance(diarization_result, str) and diarization_result.startswith("错误"):
        return diarization_result  # 出错时返回错误信息
    else:
        # 获取拼接后的音频长度
        final_audio_length = len(AudioSegment.from_wav("final_output.wav")) / 1000  # 秒为单位
        
        # 查找最匹配的说话人
        best_match, overlap_duration = find_best_matching_speaker(
            time_dict['start_time'], 
            time_dict['end_time'], 
            diarization_result
        )
        
        if best_match:
            # 获取目标说话人的时间段（排除和截断目标音频时间段）
            speaker_segments = get_speaker_segments(
                diarization_result, 
                time_dict['start_time'], 
                time_dict['end_time'], 
                final_audio_length
            )
            
            if best_match in speaker_segments:
                # 拼接所有片段
                final_output = AudioSegment.empty()
                for segment in speaker_segments[best_match]:
                    start_time_ms = int(segment[0] * 1000)  # 转为毫秒
                    end_time_ms = int(segment[1] * 1000)
                    segment_audio = AudioSegment.from_wav("final_output.wav")[start_time_ms:end_time_ms]
                    final_output += segment_audio
                
                # 导出最终拼接音频
                final_output.export("final_combined_output.wav", format="wav")
                
                return "final_combined_output.wav"
            else:
                return "没有找到匹配的说话人时间段。"
        else:
            return "未找到匹配的说话人。"

# Gradio 接口
with gr.Blocks() as demo:
    gr.Markdown("""
    # 🗣️ 音频拼接与说话人分类 🗣️
    上传目标音频和混合音频，拼接并进行说话人分类。
    结果包括目标说话人（SPEAKER_00）的时间段，已排除和截断目标录音时间段，并自动剪辑目标音频。
    """)
    
    mixed_audio_input = gr.Audio(type="filepath", label="上传混合音频")
    target_audio_input = gr.Audio(type="filepath", label="上传目标说话人音频")
    
    process_button = gr.Button("处理音频")
    
    # 输出结果
    output_audio = gr.Audio(label="剪辑后的音频")

    # 点击按钮时触发处理音频
    process_button.click(
        fn=process_audio,
        inputs=[target_audio_input, mixed_audio_input],
        outputs=[output_audio]
    )

demo.launch(share=True)