app.py

import gradio as gr
import librosa
import numpy as np
import torch
from transformers import pipeline
from tqdm import tqdm
import tempfile
import json
import soundfile as sf

# ==== 参数设置 ====
SAMPLE_RATE = 8000        # 降采样，节省计算
WINDOW = 5                 # 每个分析窗口长度（秒）
STEP = 2                   # 滑动步长（秒）
MUSIC_THRESHOLD = 0.4
VOICE_THRESHOLD = 0.3
MIN_SING_DURATION = 8      # 最短唱歌片段（秒）

# ==== 初始化模型 ====
music_pipe = pipeline("audio-classification", model="AI-Music-Detection/ai_music_detection_large_60s")
voice_pipe = pipeline("audio-classification", model="superb/hubert-large-superb-sid")

def detect_singing(audio_path):
    """核心：检测唱歌时间戳"""
    wav, sr = librosa.load(audio_path, sr=SAMPLE_RATE)
    duration = len(wav) / SAMPLE_RATE
    results = []

    for start in np.arange(0, duration - WINDOW, STEP):
        end = start + WINDOW
        snippet = wav[int(start * SAMPLE_RATE):int(end * SAMPLE_RATE)]

        # 音乐概率
        music_pred = music_pipe(snippet, sampling_rate=SAMPLE_RATE)
        music_score = max([p['score'] for p in music_pred if 'music' in p['label'].lower()] or [0])

        # 声音概率（有语音活动）
        voice_pred = voice_pipe(snippet, sampling_rate=SAMPLE_RATE)
        voice_score = max([p['score'] for p in voice_pred if 'speech' in p['label'].lower()] or [0])

        if music_score > MUSIC_THRESHOLD and voice_score > VOICE_THRESHOLD:
            results.append((float(start), float(end)))

    # 合并连续区间
    merged = []
    for seg in results:
        if not merged or seg[0] > merged[-1][1]:
            merged.append(list(seg))
        else:
            merged[-1][1] = seg[1]

    merged = [(s, e) for s, e in merged if e - s >= MIN_SING_DURATION]

    return merged


def analyze_audio(file):
    """Gradio 接口函数"""
    if file is None:
        return "请上传音频文件", None

    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
        # 确保为wav格式
        data, sr = librosa.load(file.name, sr=SAMPLE_RATE)
        sf.write(tmp.name, data, sr)
        segments = detect_singing(tmp.name)

    if not segments:
        return "未检测到明显唱歌片段", json.dumps([], indent=2)

    json_output = json.dumps(
        [{"start": s, "end": e, "duration": round(e - s, 2)} for s, e in segments],
        indent=2
    )
    return f"检测到 {len(segments)} 段唱歌片段", json_output


# ==== Gradio UI ====
with gr.Blocks(title="🎵 Singing Segment Detector") as demo:
    gr.Markdown("# 🎤 自动识别唱歌片段 (Hugging Face Space)\n上传音频文件，返回检测到的唱歌时间段 JSON。")
    audio_in = gr.Audio(type="filepath", label="上传音频文件（从视频提取后）")
    btn = gr.Button("开始分析")
    status = gr.Textbox(label="分析结果", interactive=False)
    json_out = gr.Code(label="唱歌片段时间戳（JSON）", language="json")
    btn.click(fn=analyze_audio, inputs=[audio_in], outputs=[status, json_out])

demo.launch()