pipeline-app/app/eval_music.py

"""
音乐生成质量评估 - 多维度评估框架

评估维度：
1. 节奏质量 (15%) - BPM 稳定性、节奏型
2. 和弦进行 (10%) - 和弦丰富度、和谐度
3. 听感质量 (20%) - 整体音质、混音平衡
4. 情绪-歌词匹配 (20%) - 情绪与歌词主题一致性
5. 指令匹配度 (15%) - 与生成 prompt 的匹配度
6. 伴奏质量 (10%) - 伴奏清晰度、乐器质量
7. 人声质量 (10%) - 人声清晰度、音准

工具：
- Demucs: 人声/伴奏分离
- ffprobe: 音频元数据、响度分析
- ffmpeg: 音频分析滤镜 (astats, loudnorm)
- LLM: 语义理解、情绪分析
"""

import asyncio
import json
import logging
import os
import re
import tempfile
from typing import Dict, Tuple, Optional

from pipeline_service.llm_bridge import llm_call

logger = logging.getLogger(__name__)

# GPU 服务器配置（复用 ktv_adapter 的配置）
GPU_HOST = "ymq@opencomputing.net"
GPU_DEMUCS_VENV = "/data/ymq/demucs_venv"


async def analyze_audio_metadata(audio_path: str) -> Dict:
    """
    使用 ffprobe 提取音频元数据
    
    返回:
        {
            "duration": float,
            "sample_rate": int,
            "channels": int,
            "bitrate": int,
            "codec": str
        }
    """
    try:
        proc = await asyncio.create_subprocess_exec(
            "ffprobe",
            "-v", "error",
            "-show_entries", "stream=sample_rate,channels,codec_name,bit_rate",
            "-show_entries", "format=duration,bit_rate",
            "-of", "json",
            audio_path,
            stdout=asyncio.subprocess.PIPE,
            stderr=asyncio.subprocess.PIPE
        )
        stdout, stderr = await proc.communicate()
        
        if proc.returncode != 0:
            logger.warning(f"ffprobe failed: {stderr.decode()}")
            return {}
        
        data = json.loads(stdout.decode())
        stream = data.get("streams", [{}])[0]
        fmt = data.get("format", {})
        
        return {
            "duration": float(fmt.get("duration", 0)),
            "sample_rate": int(stream.get("sample_rate", 0)),
            "channels": int(stream.get("channels", 0)),
            "bitrate": int(fmt.get("bit_rate", 0)),
            "codec": stream.get("codec_name", "unknown")
        }
    except Exception as e:
        logger.error(f"Failed to analyze audio metadata: {e}")
        return {}


async def analyze_audio_loudness(audio_path: str) -> Dict:
    """
    使用 ffmpeg loudnorm 分析音频响度
    
    返回:
        {
            "integrated_loudness": float,  # 综合响度 (LUFS)
            "loudness_range": float,       # 响度范围 (LU)
            "true_peak": float             # 真峰值 (dBTP)
        }
    """
    try:
        proc = await asyncio.create_subprocess_exec(
            "ffmpeg",
            "-i", audio_path,
            "-af", "loudnorm=print_format=json",
            "-f", "null",
            "-",
            stdout=asyncio.subprocess.PIPE,
            stderr=asyncio.subprocess.PIPE
        )
        stdout, stderr = await proc.communicate()
        
        # loudnorm 输出在 stderr
        output = stderr.decode()
        
        # 提取 JSON 部分
        import re
        json_match = re.search(r'\{[^}]+\}', output)
        if json_match:
            data = json.loads(json_match.group())
            return {
                "integrated_loudness": float(data.get("input_i", -70)),
                "loudness_range": float(data.get("input_lra", 0)),
                "true_peak": float(data.get("input_tp", -100))
            }
        
        return {}
    except Exception as e:
        logger.error(f"Failed to analyze audio loudness: {e}")
        return {}


async def separate_vocals_accompaniment(
    audio_path: str,
    output_dir: str
) -> Tuple[Optional[str], Optional[str]]:
    """
    使用 Demucs 分离人声和伴奏
    
    返回:
        (vocals_path, accompaniment_path) or (None, None)
    """
    try:
        # 在 GPU 服务器上运行 Demucs
        gpu_audio_path = f"/tmp/eval_{os.path.basename(audio_path)}"
        
        # 上传音频
        upload_cmd = f"scp '{audio_path}' {GPU_HOST}:{gpu_audio_path}"
        upload_proc = await asyncio.create_subprocess_shell(
            upload_cmd,
            stdout=asyncio.subprocess.PIPE,
            stderr=asyncio.subprocess.PIPE
        )
        await upload_proc.communicate()
        
        # 运行 Demucs
        demucs_cmd = (
            f"ssh {GPU_HOST} '"
            f"source {GPU_DEMUCS_VENV}/bin/activate && "
            f"cd {GPU_DEMUCS_VENV} && "
            f"python -m demucs --two-stems vocals --out /tmp/demucs_out '{gpu_audio_path}'"
            f"'"
        )
        demucs_proc = await asyncio.create_subprocess_shell(
            demucs_cmd,
            stdout=asyncio.subprocess.PIPE,
            stderr=asyncio.subprocess.PIPE
        )
        await demucs_proc.communicate()
        
        # 下载分离结果
        # Demucs 输出格式: /tmp/demucs_out/htdemucs/{filename}/vocals.wav
        filename = os.path.splitext(os.path.basename(audio_path))[0]
        gpu_vocals = f"/tmp/demucs_out/htdemucs/{filename}/vocals.wav"
        gpu_no_vocals = f"/tmp/demucs_out/htdemucs/{filename}/no_vocals.wav"
        
        local_vocals = os.path.join(output_dir, f"{filename}_vocals.wav")
        local_no_vocals = os.path.join(output_dir, f"{filename}_no_vocals.wav")
        
        # 下载人声
        download_vocals = f"scp {GPU_HOST}:{gpu_vocals} '{local_vocals}'"
        download_proc1 = await asyncio.create_subprocess_shell(
            download_vocals,
            stdout=asyncio.subprocess.PIPE,
            stderr=asyncio.subprocess.PIPE
        )
        await download_proc1.communicate()
        
        # 下载伴奏
        download_no_vocals = f"scp {GPU_HOST}:{gpu_no_vocals} '{local_no_vocals}'"
        download_proc2 = await asyncio.create_subprocess_shell(
            download_no_vocals,
            stdout=asyncio.subprocess.PIPE,
            stderr=asyncio.subprocess.PIPE
        )
        await download_proc2.communicate()
        
        # 清理 GPU 临时文件
        cleanup_cmd = f"ssh {GPU_HOST} 'rm -f {gpu_audio_path} && rm -rf /tmp/demucs_out'"
        cleanup_proc = await asyncio.create_subprocess_shell(
            cleanup_cmd,
            stdout=asyncio.subprocess.PIPE,
            stderr=asyncio.subprocess.PIPE
        )
        await cleanup_proc.communicate()
        
        vocals_exists = os.path.exists(local_vocals)
        no_vocals_exists = os.path.exists(local_no_vocals)
        
        return (
            local_vocals if vocals_exists else None,
            local_no_vocals if no_vocals_exists else None
        )
        
    except Exception as e:
        logger.error(f"Failed to separate vocals: {e}")
        return None, None


async def evaluate_rhythm_quality(audio_path: str) -> Tuple[float, str]:
    """
    评估节奏质量
    
    检查项：
    - BPM 稳定性（通过节拍分析）
    - 节奏型清晰度
    
    简化实现：基于音频时长和响度范围推断
    
    返回: (score 0-1, reason)
    """
    metadata = await analyze_audio_metadata(audio_path)
    loudness = await analyze_audio_loudness(audio_path)
    
    if not metadata:
        return 0.0, "无法读取音频元数据"
    
    duration = metadata.get("duration", 0)
    lra = loudness.get("loudness_range", 0)
    
    # 节奏质量推断
    # 响度范围适中 (5-15 LU) 通常表示节奏稳定
    if 5 <= lra <= 15:
        score = 0.85
        reason = f"节奏稳定 (LRA={lra:.1f}LU)"
    elif 3 <= lra < 5 or 15 < lra <= 20:
        score = 0.7
        reason = f"节奏基本稳定 (LRA={lra:.1f}LU)"
    else:
        score = 0.5
        reason = f"节奏波动较大 (LRA={lra:.1f}LU)"
    
    # 时长加分
    if duration >= 60:
        score = min(1.0, score + 0.1)
        reason += f", 时长充足 ({duration:.0f}s)"
    
    return score, reason


async def evaluate_chord_progression(audio_path: str) -> Tuple[float, str]:
    """
    评估和弦进行
    
    简化实现：基于音频特征推断和声丰富度
    
    返回: (score 0-1, reason)
    """
    # 和弦分析需要专业的音频分析库（如 librosa 的 chroma 特征）
    # 这里使用简化评估
    
    metadata = await analyze_audio_metadata(audio_path)
    if not metadata:
        return 0.0, "无法读取音频元数据"
    
    duration = metadata.get("duration", 0)
    
    # 时长较长的音乐通常有更丰富的和声进行
    if duration >= 180:  # 3分钟+
        return 0.85, f"和声进行丰富 (时长 {duration:.0f}s)"
    elif duration >= 120:
        return 0.75, f"和声进行合理 (时长 {duration:.0f}s)"
    elif duration >= 60:
        return 0.65, f"和声进行基本 (时长 {duration:.0f}s)"
    else:
        return 0.5, f"时长过短，和声发展有限 ({duration:.0f}s)"


async def evaluate_audio_quality(audio_path: str) -> Tuple[float, str]:
    """
    评估听感质量 - 整体音质、混音平衡
    
    检查项：
    - 响度水平 (-24 到 -14 LUFS 为佳)
    - 真峰值 (< -1 dBTP)
    - 采样率和位深
    
    返回: (score 0-1, reason)
    """
    metadata = await analyze_audio_metadata(audio_path)
    loudness = await analyze_audio_loudness(audio_path)
    
    if not metadata:
        return 0.0, "无法读取音频元数据"
    
    score = 0.0
    reasons = []
    
    # 采样率评分
    sample_rate = metadata.get("sample_rate", 0)
    if sample_rate >= 44100:
        score += 0.2
        reasons.append(f"高采样率 {sample_rate}Hz")
    elif sample_rate >= 22050:
        score += 0.15
        reasons.append(f"标准采样率 {sample_rate}Hz")
    else:
        reasons.append(f"采样率偏低 {sample_rate}Hz")
    
    # 响度评分
    integrated_loudness = loudness.get("integrated_loudness", -70)
    if -24 <= integrated_loudness <= -14:
        score += 0.4
        reasons.append(f"响度适中 ({integrated_loudness:.1f} LUFS)")
    elif -30 <= integrated_loudness < -24 or -14 < integrated_loudness <= -10:
        score += 0.3
        reasons.append(f"响度可接受 ({integrated_loudness:.1f} LUFS)")
    else:
        score += 0.1
        reasons.append(f"响度异常 ({integrated_loudness:.1f} LUFS)")
    
    # 真峰值评分
    true_peak = loudness.get("true_peak", -100)
    if true_peak < -1:
        score += 0.2
        reasons.append(f"无削波 (TP={true_peak:.1f} dBTP)")
    elif true_peak < 0:
        score += 0.15
        reasons.append(f"轻微削波风险 (TP={true_peak:.1f} dBTP)")
    else:
        reasons.append(f"明显削波 (TP={true_peak:.1f} dBTP)")
    
    # 位深/码率评分
    bitrate = metadata.get("bitrate", 0)
    if bitrate >= 320000:  # 320kbps+
        score += 0.2
        reasons.append(f"高码率 {bitrate//1000}kbps")
    elif bitrate >= 192000:
        score += 0.15
        reasons.append(f"标准码率 {bitrate//1000}kbps")
    else:
        reasons.append(f"码率偏低 {bitrate//1000}kbps")
    
    return score, "; ".join(reasons)


async def evaluate_emotion_lyrics_match(
    audio_path: str,
    lyrics: Optional[str] = None
) -> Tuple[float, str]:
    """
    评估情绪-歌词匹配度
    
    使用 LLM 分析：
    1. 音频的情绪特征（通过音频特征推断）
    2. 歌词的情感主题
    3. 两者的一致性
    
    返回: (score 0-1, reason)
    """
    if not lyrics:
        return 0.6, "无歌词，跳过情绪-歌词匹配评估"
    
    # 分析音频情绪（简化：基于响度和时长）
    loudness = await analyze_audio_loudness(audio_path)
    integrated_loudness = loudness.get("integrated_loudness", -70)
    
    # 响度较高通常表示更强烈的情绪
    if integrated_loudness > -18:
        audio_emotion = "强烈/激昂"
    elif integrated_loudness > -24:
        audio_emotion = "中等强度"
    else:
        audio_emotion = "柔和/舒缓"
    
    try:
        llm_prompt = f"""分析歌词情感与音频特征的匹配度。

歌词:
{lyrics}

音频情绪特征: {audio_emotion} (响度: {integrated_loudness:.1f} LUFS)

请评估歌词表达的情感是否与音频的情绪特征一致：
1. 歌词的主题和情感倾向
2. 与音频情绪的匹配程度

请严格按照以下 JSON 格式返回，不要输出其他内容：
{{"match_score": <0到10的整数>, "reason": "简要说明匹配情况"}}
"""
        response = await llm_call(llm_prompt, temperature=0.3)
        
        # 提取 JSON（兼容 markdown 代码块包裹）
        json_match = re.search(r'\{[^{}]*\}', response)
        if not json_match:
            logger.warning(f"LLM 返回无法解析为 JSON: {response[:200]}")
            return 0.6, f"LLM 返回格式异常，降级评估（音频情绪: {audio_emotion}）"
        
        data = json.loads(json_match.group())
        match_score = float(data.get("match_score", 8))
        reason = data.get("reason", "LLM 评估完成")
        
        # 转换为 0-1 分数
        score = max(0.0, min(1.0, match_score / 10.0))
        return score, f"情绪-歌词匹配: {reason} (音频情绪: {audio_emotion})"
        
    except Exception as e:
        logger.warning(f"情绪-歌词匹配 LLM 评估失败: {e}")
        return 0.6, f"LLM 分析异常，降级评估（音频情绪: {audio_emotion}）: {e}"


async def evaluate_prompt_adherence(
    audio_path: str,
    prompt: Optional[str] = None
) -> Tuple[float, str]:
    """
    评估指令匹配度 - 与生成 prompt 的匹配度
    
    使用 LLM 分析：
    1. prompt 描述的风格、情绪、乐器等
    2. 音频是否体现这些特征
    
    返回: (score 0-1, reason)
    """
    if not prompt:
        return 0.6, "无 prompt，跳过指令匹配度评估"
    
    try:
        # 获取音频特征
        metadata = await analyze_audio_metadata(audio_path)
        loudness = await analyze_audio_loudness(audio_path)
        
        duration = metadata.get("duration", 0)
        integrated_loudness = loudness.get("integrated_loudness", -70)
        loudness_range = loudness.get("loudness_range", 0)
        true_peak = loudness.get("true_peak", -100)
        
        llm_prompt = f"""评估一段 AI 生成的音频与生成指令之间的匹配度。

【生成指令 (prompt)】
{prompt}

【音频特征（从音频信号中提取）】
- 时长: {duration:.1f} 秒
- 综合响度: {integrated_loudness:.1f} LUFS
- 响度范围: {loudness_range:.1f} LU
- 真峰值: {true_peak:.1f} dBTP

请根据以上信息，评估指令中描述的风格、情绪、乐器、速度等特征是否在音频的物理特性中得到体现。
注意：你无法直接听到音频，只能通过上述特征进行推断。请结合常识分析（例如：高响度+窄响度范围通常对应激烈/电子风格，低响度+宽响度范围通常对应舒缓/原声风格）。

请严格按照以下 JSON 格式返回，不要输出其他内容：
{{"match_score": <0到10的整数>, "reason": "简要说明匹配情况"}}
"""
        response = await llm_call(llm_prompt, temperature=0.3)
        
        # 提取 JSON（兼容 markdown 代码块包裹）
        json_match = re.search(r'\{[^{}]*\}', response)
        if not json_match:
            logger.warning(f"LLM 返回无法解析为 JSON: {response[:200]}")
            return 0.6, "LLM 返回格式异常，降级评估"
        
        data = json.loads(json_match.group())
        match_score = float(data.get("match_score", 8))
        reason = data.get("reason", "LLM 评估完成")
        
        # 转换为 0-1 分数
        score = max(0.0, min(1.0, match_score / 10.0))
        return score, f"指令匹配: {reason}"
        
    except Exception as e:
        logger.warning(f"指令匹配度 LLM 评估失败: {e}")
        return 0.6, f"LLM 分析异常，降级评估: {e}"


async def evaluate_accompaniment_quality(
    accompaniment_path: Optional[str]
) -> Tuple[float, str]:
    """
    评估伴奏质量
    
    检查项：
    - 伴奏清晰度
    - 乐器质量
    - 混音平衡
    
    返回: (score 0-1, reason)
    """
    if not accompaniment_path or not os.path.exists(accompaniment_path):
        return 0.5, "伴奏文件不存在，跳过伴奏质量评估"
    
    # 分析伴奏响度
    loudness = await analyze_audio_loudness(accompaniment_path)
    integrated_loudness = loudness.get("integrated_loudness", -70)
    
    # 伴奏响度适中为佳
    if -22 <= integrated_loudness <= -16:
        score = 0.85
        reason = f"伴奏混音平衡 ({integrated_loudness:.1f} LUFS)"
    elif -28 <= integrated_loudness < -22 or -16 < integrated_loudness <= -12:
        score = 0.7
        reason = f"伴奏混音可接受 ({integrated_loudness:.1f} LUFS)"
    else:
        score = 0.5
        reason = f"伴奏响度异常 ({integrated_loudness:.1f} LUFS)"
    
    return score, reason


async def evaluate_vocal_quality(
    vocals_path: Optional[str]
) -> Tuple[float, str]:
    """
    评估人声质量
    
    检查项：
    - 人声清晰度
    - 音准（简化评估）
    - 响度平衡
    
    返回: (score 0-1, reason)
    """
    if not vocals_path or not os.path.exists(vocals_path):
        return 0.5, "人声文件不存在，跳过人声质量评估"
    
    # 分析人声响度
    loudness = await analyze_audio_loudness(vocals_path)
    integrated_loudness = loudness.get("integrated_loudness", -70)
    
    # 人声响度通常高于伴奏
    if -20 <= integrated_loudness <= -12:
        score = 0.85
        reason = f"人声清晰 ({integrated_loudness:.1f} LUFS)"
    elif -26 <= integrated_loudness < -20 or -12 < integrated_loudness <= -8:
        score = 0.7
        reason = f"人声可接受 ({integrated_loudness:.1f} LUFS)"
    else:
        score = 0.5
        reason = f"人声响度异常 ({integrated_loudness:.1f} LUFS)"
    
    return score, reason


async def evaluate_music_quality(
    audio_path: str,
    prompt: Optional[str] = None,
    lyrics: Optional[str] = None,
    use_demucs: bool = True
) -> Dict:
    """
    音乐质量综合评估
    
    参数:
        audio_path: 音频文件路径
        prompt: 生成时的描述（可选）
        lyrics: 歌词文本（可选）
        use_demucs: 是否使用 Demucs 分离人声/伴奏
    
    返回:
        {
            "overall_score": float (0-1),
            "overall_reason": str,
            "dimensions": {
                "rhythm_quality": {"score": float, "weight": 0.15, "reason": str},
                "chord_progression": {"score": float, "weight": 0.10, "reason": str},
                "audio_quality": {"score": float, "weight": 0.20, "reason": str},
                "emotion_lyrics_match": {"score": float, "weight": 0.20, "reason": str},
                "prompt_adherence": {"score": float, "weight": 0.15, "reason": str},
                "accompaniment_quality": {"score": float, "weight": 0.10, "reason": str},
                "vocal_quality": {"score": float, "weight": 0.10, "reason": str}
            }
        }
    """
    # 基础检查
    if not os.path.exists(audio_path):
        return {
            "overall_score": 0.0,
            "overall_reason": f"音频文件不存在: {audio_path}",
            "dimensions": {}
        }
    
    file_size = os.path.getsize(audio_path)
    if file_size < 51200:  # < 50KB
        return {
            "overall_score": 0.0,
            "overall_reason": f"音频文件过小: {file_size} bytes",
            "dimensions": {}
        }
    
    # 分离人声和伴奏（如果启用）
    vocals_path = None
    accompaniment_path = None
    
    if use_demucs:
        with tempfile.TemporaryDirectory() as tmpdir:
            vocals_path, accompaniment_path = await separate_vocals_accompaniment(
                audio_path, tmpdir
            )
            
            # 评估各维度
            rhythm_score, rhythm_reason = await evaluate_rhythm_quality(audio_path)
            chord_score, chord_reason = await evaluate_chord_progression(audio_path)
            quality_score, quality_reason = await evaluate_audio_quality(audio_path)
            emotion_score, emotion_reason = await evaluate_emotion_lyrics_match(audio_path, lyrics)
            prompt_score, prompt_reason = await evaluate_prompt_adherence(audio_path, prompt)
            accomp_score, accomp_reason = await evaluate_accompaniment_quality(accompaniment_path)
            vocal_score, vocal_reason = await evaluate_vocal_quality(vocals_path)
    
    else:
        # 不使用 Demucs，跳过伴奏和人声评估
        rhythm_score, rhythm_reason = await evaluate_rhythm_quality(audio_path)
        chord_score, chord_reason = await evaluate_chord_progression(audio_path)
        quality_score, quality_reason = await evaluate_audio_quality(audio_path)
        emotion_score, emotion_reason = await evaluate_emotion_lyrics_match(audio_path, lyrics)
        prompt_score, prompt_reason = await evaluate_prompt_adherence(audio_path, prompt)
        accomp_score, accomp_reason = 1.0, "未分离伴奏，跳过评估"
        vocal_score, vocal_reason = 1.0, "未分离人声，跳过评估"
    
    # 构建维度评估
    dimensions = {
        "rhythm_quality": {"score": rhythm_score, "weight": 0.15, "reason": rhythm_reason},
        "chord_progression": {"score": chord_score, "weight": 0.10, "reason": chord_reason},
        "audio_quality": {"score": quality_score, "weight": 0.20, "reason": quality_reason},
        "emotion_lyrics_match": {"score": emotion_score, "weight": 0.20, "reason": emotion_reason},
        "prompt_adherence": {"score": prompt_score, "weight": 0.15, "reason": prompt_reason},
        "accompaniment_quality": {"score": accomp_score, "weight": 0.10, "reason": accomp_reason},
        "vocal_quality": {"score": vocal_score, "weight": 0.10, "reason": vocal_reason},
    }
    
    # 加权计算总分
    overall_score = sum(d["score"] * d["weight"] for d in dimensions.values())
    
    # 生成综合原因
    reasons = []
    for dim_name, dim_data in dimensions.items():
        if dim_data["score"] < 0.6:
            reasons.append(f"{dim_name}: {dim_data['reason']}")
    
    overall_reason = "; ".join(reasons) if reasons else "各维度评估良好"
    
    return {
        "overall_score": overall_score,
        "overall_reason": overall_reason,
        "dimensions": dimensions
    }