app.py

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

# 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

# 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}")

    # 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
@spaces.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)
    except Exception as e:
        return f"处理音频时出错: {e}"
    print(diarization)  
    print(type(diarization))
    # 返回 diarization 输出
    return str(diarization)

# 将时间戳转换为秒
def timestamp_to_seconds(timestamp):
    try:
        h, m, s = map(float, timestamp.split(':'))
        return 3600 * h + 60 * m + s
    except ValueError as e:
        print(f"转换时间戳时出错: '{timestamp}'. 错误: {e}")
        return None

# 计算时间段的重叠部分（单位：秒）
def calculate_overlap(start1, end1, start2, end2):
    overlap_start = max(start1, start2)
    overlap_end = min(end1, end2)
    overlap_duration = max(0, overlap_end - overlap_start)
    return overlap_duration

# 获取目标时间段和说话人时间段的重叠比例
def get_best_match(target_time, diarization_output):
    target_start_time = target_time['start_time']
    target_end_time = target_time['end_time']
    
    # 假设 diarization_output 是一个列表，包含说话人时间段和标签
    speaker_segments = []
    for line in diarization_output.strip().split('\n'):
        try:
            parts = line.strip()[1:-1].split(' --> ')
            start_time = parts[0].strip()
            end_time = parts[1].split(']')[0].strip()
            label = line.split()[-1].strip()

            start_seconds = timestamp_to_seconds(start_time)
            end_seconds = timestamp_to_seconds(end_time)

            # 计算目标音频时间段和说话人时间段的重叠时间
            overlap = calculate_overlap(target_start_time, target_end_time, start_seconds, end_seconds)
            overlap_ratio = overlap / (end_seconds - start_seconds)

            # 记录说话人标签和重叠比例
            speaker_segments.append((label, overlap_ratio, start_seconds, end_seconds))

        except Exception as e:
            print(f"处理行时出错: '{line.strip()}'. 错误: {e}")
    
    # 按照重叠比例排序，返回重叠比例最大的一段
    best_match = max(speaker_segments, key=lambda x: x[1], default=None)
    
    return best_match

# 处理音频文件并返回输出
def process_audio(target_audio, mixed_audio):
    # 打印文件路径，确保传入的文件有效
    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)
    
    # 执行说话人分离
    diarization_result = diarize_audio("final_output.wav")
    
    if diarization_result.startswith("错误"):
        return diarization_result, None  # 出错时返回错误信息
    else:
        # 获取最佳匹配的说话人标签和时间段
        best_match = get_best_match(time_dict, diarization_result)
        
        if best_match:
            # 返回最佳匹配说话人的标签和时间段
            return best_match[0], best_match[2], best_match[3]

# 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。结果包括最佳匹配说话人的时间段。
    """)
    
    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
    diarization_output = gr.Textbox(label="最佳匹配说话人")
    time_range_output = gr.Textbox(label="最佳匹配时间段")

    # 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=[diarization_output, time_range_output]
    )

demo.launch(share=True)