neverlib 0.2.2__py3-none-any.whl → 0.2.4__py3-none-any.whl

neverlib/data_analyze/quality_metrics.py

@@ -0,0 +1,364 @@
+"""
+音频质量评估模块
+Audio Quality Metrics Module
+
+提供音频质量评估和失真度分析功能
+"""
+
+import numpy as np
+import librosa
+from scipy import signal
+from scipy.fft import fft, fftfreq
+from typing import Tuple, Optional, Union, List
+import warnings
+
+
+class QualityAnalyzer:
+    """音频质量分析器类"""
+    
+    def __init__(self, sr: int = 22050):
+        """
+        初始化质量分析器
+        
+        Args:
+            sr: 采样率
+        """
+        self.sr = sr
+    
+    def signal_to_noise_ratio(self, signal_audio: np.ndarray, 
+                             noise_audio: Optional[np.ndarray] = None,
+                             signal_start: Optional[int] = None,
+                             signal_end: Optional[int] = None) -> float:
+        """
+        计算信噪比 (SNR)
+        
+        Args:
+            signal_audio: 含有信号和噪声的音频
+            noise_audio: 纯噪声音频（可选）
+            signal_start: 信号开始位置（当噪声未单独提供时使用）
+            signal_end: 信号结束位置（当噪声未单独提供时使用）
+            
+        Returns:
+            SNR值（dB）
+        """
+        if noise_audio is not None:
+            # 如果提供了噪声音频
+            signal_power = np.mean(signal_audio ** 2)
+            noise_power = np.mean(noise_audio ** 2)
+        else:
+            # 从音频中提取信号和噪声部分
+            if signal_start is None or signal_end is None:
+                raise ValueError("Must provide signal_start and signal_end when noise_audio is None")
+            
+            signal_part = signal_audio[signal_start:signal_end]
+            
+            # 假设开头和结尾是噪声
+            noise_start = signal_audio[:signal_start] if signal_start > 0 else np.array([])
+            noise_end = signal_audio[signal_end:] if signal_end < len(signal_audio) else np.array([])
+            noise_part = np.concatenate([noise_start, noise_end]) if len(noise_start) > 0 or len(noise_end) > 0 else signal_audio[:1000]
+            
+            signal_power = np.mean(signal_part ** 2)
+            noise_power = np.mean(noise_part ** 2)
+        
+        if noise_power == 0:
+            return float('inf')
+        
+        snr_db = 10 * np.log10(signal_power / noise_power)
+        return snr_db
+    
+    def total_harmonic_distortion(self, audio: np.ndarray, 
+                                 fundamental_freq: Optional[float] = None,
+                                 num_harmonics: int = 5) -> float:
+        """
+        计算总谐波失真 (THD)
+        
+        Args:
+            audio: 音频信号
+            fundamental_freq: 基频（Hz）, 如果不提供则自动检测
+            num_harmonics: 考虑的谐波数量
+            
+        Returns:
+            THD百分比
+        """
+        # 计算频谱
+        spectrum = fft(audio)
+        freqs = fftfreq(len(audio), 1/self.sr)
+        magnitude = np.abs(spectrum)
+        
+        # 只考虑正频率
+        positive_idx = freqs > 0
+        freqs = freqs[positive_idx]
+        magnitude = magnitude[positive_idx]
+        
+        # 如果没有提供基频, 自动检测
+        if fundamental_freq is None:
+            fundamental_freq = freqs[np.argmax(magnitude)]
+        
+        # 找到基频和谐波的功率
+        tolerance = fundamental_freq * 0.05  # 5%的容差
+        
+        # 基频功率
+        fundamental_idx = np.where(np.abs(freqs - fundamental_freq) < tolerance)[0]
+        if len(fundamental_idx) == 0:
+            return 0.0
+        
+        fundamental_power = np.max(magnitude[fundamental_idx]) ** 2
+        
+        # 谐波功率
+        harmonic_power = 0
+        for h in range(2, num_harmonics + 2):
+            harmonic_freq = h * fundamental_freq
+            harmonic_idx = np.where(np.abs(freqs - harmonic_freq) < tolerance)[0]
+            if len(harmonic_idx) > 0:
+                harmonic_power += np.max(magnitude[harmonic_idx]) ** 2
+        
+        if fundamental_power == 0:
+            return 0.0
+        
+        thd = np.sqrt(harmonic_power / fundamental_power) * 100
+        return thd
+    
+    def dynamic_range(self, audio: np.ndarray, percentile_low: float = 1, 
+                     percentile_high: float = 99) -> float:
+        """
+        计算动态范围
+        
+        Args:
+            audio: 音频信号
+            percentile_low: 低百分位数
+            percentile_high: 高百分位数
+            
+        Returns:
+            动态范围（dB）
+        """
+        amplitude = np.abs(audio)
+        amplitude = amplitude[amplitude > 0]  # 避免log(0)
+        
+        if len(amplitude) == 0:
+            return 0.0
+        
+        low_level = np.percentile(amplitude, percentile_low)
+        high_level = np.percentile(amplitude, percentile_high)
+        
+        dynamic_range_db = 20 * np.log10(high_level / (low_level + 1e-10))
+        return dynamic_range_db
+    
+    def frequency_response(self, audio: np.ndarray, 
+                          reference_audio: Optional[np.ndarray] = None) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        计算频率响应特性
+        
+        Args:
+            audio: 测试音频信号
+            reference_audio: 参考音频信号（可选）
+            
+        Returns:
+            (频率数组, 幅度响应数组)
+        """
+        if reference_audio is not None:
+            # 计算传递函数
+            freqs, h = signal.freqz_zpk(*signal.tf2zpk([1], [1]), fs=self.sr)
+            
+            # 使用互相关计算频率响应
+            cross_corr = signal.correlate(audio, reference_audio, mode='full')
+            auto_corr = signal.correlate(reference_audio, reference_audio, mode='full')
+            
+            # 频域除法得到传递函数
+            cross_spectrum = fft(cross_corr)
+            auto_spectrum = fft(auto_corr)
+            
+            h_measured = cross_spectrum / (auto_spectrum + 1e-10)
+            freqs = fftfreq(len(h_measured), 1/self.sr)
+            
+            # 只取正频率部分
+            positive_idx = freqs >= 0
+            freqs = freqs[positive_idx]
+            h_measured = h_measured[positive_idx]
+            
+            return freqs, np.abs(h_measured)
+        else:
+            # 直接返回频谱
+            spectrum = fft(audio)
+            freqs = fftfreq(len(audio), 1/self.sr)
+            
+            positive_idx = freqs >= 0
+            freqs = freqs[positive_idx]
+            spectrum = spectrum[positive_idx]
+            
+            return freqs, np.abs(spectrum)
+    
+    def loudness_range(self, audio: np.ndarray, gate_threshold: float = -70) -> dict:
+        """
+        计算响度范围（基于EBU R128标准的简化版本）
+        
+        Args:
+            audio: 音频信号
+            gate_threshold: 门限阈值（dB）
+            
+        Returns:
+            响度统计信息字典
+        """
+        # 分块计算短时响度
+        block_size = int(0.4 * self.sr)  # 400ms块
+        hop_size = int(0.1 * self.sr)    # 100ms跳跃
+        
+        blocks = []
+        for i in range(0, len(audio) - block_size, hop_size):
+            block = audio[i:i + block_size]
+            # 简化的响度计算（使用RMS近似）
+            rms = np.sqrt(np.mean(block ** 2))
+            if rms > 0:
+                loudness = 20 * np.log10(rms)
+                if loudness > gate_threshold:
+                    blocks.append(loudness)
+        
+        if len(blocks) == 0:
+            return {'integrated_loudness': -float('inf'), 'loudness_range': 0, 'max_loudness': -float('inf')}
+        
+        blocks = np.array(blocks)
+        
+        # 计算统计量
+        integrated_loudness = np.mean(blocks)
+        loudness_range = np.percentile(blocks, 95) - np.percentile(blocks, 10)
+        max_loudness = np.max(blocks)
+        
+        return {
+            'integrated_loudness': integrated_loudness,
+            'loudness_range': loudness_range,
+            'max_loudness': max_loudness
+        }
+    
+    def spectral_distortion(self, original: np.ndarray, processed: np.ndarray) -> float:
+        """
+        计算谱失真度
+        
+        Args:
+            original: 原始音频
+            processed: 处理后音频
+            
+        Returns:
+            谱失真度（dB）
+        """
+        # 确保两个信号长度相同
+        min_len = min(len(original), len(processed))
+        original = original[:min_len]
+        processed = processed[:min_len]
+        
+        # 计算频谱
+        orig_spectrum = np.abs(fft(original))
+        proc_spectrum = np.abs(fft(processed))
+        
+        # 计算谱失真
+        mse = np.mean((orig_spectrum - proc_spectrum) ** 2)
+        orig_power = np.mean(orig_spectrum ** 2)
+        
+        if orig_power == 0:
+            return float('inf')
+        
+        distortion_db = 10 * np.log10(mse / orig_power)
+        return distortion_db
+
+
+def comprehensive_quality_assessment(audio: np.ndarray, sr: int = 22050, 
+                                   reference: Optional[np.ndarray] = None) -> dict:
+    """
+    综合质量评估
+    
+    Args:
+        audio: 待评估音频
+        sr: 采样率
+        reference: 参考音频（可选）
+        
+    Returns:
+        质量评估结果字典
+    """
+    analyzer = QualityAnalyzer(sr=sr)
+    
+    results = {
+        'dynamic_range': analyzer.dynamic_range(audio),
+        'loudness_stats': analyzer.loudness_range(audio),
+    }
+    
+    # 尝试计算THD
+    try:
+        results['thd'] = analyzer.total_harmonic_distortion(audio)
+    except:
+        results['thd'] = None
+    
+    # 如果有参考音频, 计算比较指标
+    if reference is not None:
+        try:
+            results['snr'] = analyzer.signal_to_noise_ratio(audio, reference)
+            results['spectral_distortion'] = analyzer.spectral_distortion(reference, audio)
+        except:
+            results['snr'] = None
+            results['spectral_distortion'] = None
+    
+    # 频率响应
+    try:
+        freqs, response = analyzer.frequency_response(audio, reference)
+        results['frequency_response'] = {
+            'frequencies': freqs,
+            'magnitude': response
+        }
+    except:
+        results['frequency_response'] = None
+    
+    return results
+
+
+def audio_health_check(audio: np.ndarray, sr: int = 22050) -> dict:
+    """
+    音频健康检查
+    
+    Args:
+        audio: 音频信号
+        sr: 采样率
+        
+    Returns:
+        健康检查结果
+    """
+    health_report = {
+        'issues': [],
+        'warnings': [],
+        'stats': {}
+    }
+    
+    # 基础统计
+    max_amplitude = np.max(np.abs(audio))
+    min_amplitude = np.min(np.abs(audio))
+    mean_amplitude = np.mean(np.abs(audio))
+    
+    health_report['stats'] = {
+        'max_amplitude': max_amplitude,
+        'min_amplitude': min_amplitude,
+        'mean_amplitude': mean_amplitude,
+        'duration': len(audio) / sr
+    }
+    
+    # 检查削波
+    if max_amplitude >= 0.99:
+        health_report['issues'].append('Potential clipping detected')
+    
+    # 检查过低音量
+    if max_amplitude < 0.01:
+        health_report['warnings'].append('Very low signal level')
+    
+    # 检查静音
+    if mean_amplitude < 1e-6:
+        health_report['issues'].append('Signal appears to be silent')
+    
+    # 检查DC偏移
+    dc_offset = np.mean(audio)
+    if abs(dc_offset) > 0.01:
+        health_report['warnings'].append(f'DC offset detected: {dc_offset:.4f}')
+    
+    # 检查动态范围
+    analyzer = QualityAnalyzer(sr=sr)
+    dynamic_range = analyzer.dynamic_range(audio)
+    if dynamic_range < 6:
+        health_report['warnings'].append('Low dynamic range')
+    elif dynamic_range > 60:
+        health_report['warnings'].append('Very high dynamic range - check for noise')
+    
+    return health_report

neverlib/data_analyze/rms_distrubution.py

@@ -0,0 +1,62 @@
+'''
+Author: 凌逆战 | Never
+Date: 2025-03-26 22:13:22
+Description: 
+'''
+# -*- coding:utf-8 -*-
+# Author:凌逆战 | Never
+# Date: 2025/3/2
+"""
+统计音频语音段rms值分布
+"""
+import sys
+sys.path.append("../../../")
+import torch
+import soundfile as sf
+from neverlib.utils import get_path_list
+from neverlib.filter import HPFilter
+from neverlib.audio_aug import volume_norm
+from neverlib.dataAnalyze.utils import rms_amplitude
+from joblib import Parallel, delayed
+import matplotlib.pyplot as plt
+import numpy as np
+import librosa
+import os
+from utils.train_utils import from_path_get_vadpoint
+
+
+
+    
+def get_rms_vad(wav_path):
+    wav, wav_sr = sf.read(wav_path, always_2d=True)  # (xxx,ch)
+    assert wav_sr == sr, f"期望采样率为{sr}, 但是为{wav_sr}, 文件名: {wav_path}"
+    vadstart, vadend = from_path_get_vadpoint(wav_path)
+    rms = rms_amplitude(wav[vadstart:vadend]).mean()
+    # if rms < -75:
+    #     print(wav_path, np.round(rms, 2))
+    # if rms > -5:
+    #     print(wav_path, np.round(rms, 2))
+    return rms
+
+
+sr = 16000
+wav_dir_list = [
+    "/data/never/Dataset/kws_data/Command_Word/Crowdsourcing/en_kws2/train/RealPerson",
+    "/data/never/Dataset/kws_data/Command_Word/Crowdsourcing/en_kws2/val/RealPerson",
+    "/data/never/Dataset/kws_data/Command_Word/Crowdsourcing/en_kws2/test/RealPerson",
+]
+wav_path_list = []
+for wav_dir in wav_dir_list:
+    wav_path_list.extend(get_path_list(wav_dir, end="*.wav"))
+
+rms_list = Parallel(n_jobs=64)(delayed(get_rms_vad)(wav_path) for wav_path in wav_path_list)
+
+# 绘制时长分布直方图
+plt.hist(rms_list, bins=100, edgecolor='black')
+plt.title("RMS Distribution")
+plt.xlabel("RMS (dB)")
+plt.ylabel("number")
+plt.grid(True)
+plt.tight_layout()
+plt.savefig("./png_dist/rms_distribution.png")
+

neverlib/data_analyze/spectral_analysis.py

@@ -0,0 +1,218 @@
+"""
+频域分析模块
+Spectral Analysis Module
+
+提供音频频域特征提取和分析功能
+"""
+
+import numpy as np
+import librosa
+import scipy.signal
+from scipy.fft import fft, fftfreq
+from typing import Tuple, Optional, Union
+import warnings
+
+
+class SpectralAnalyzer:
+    """频谱分析器类"""
+    
+    def __init__(self, sr: int = 22050, n_fft: int = 2048, hop_length: int = 512):
+        """
+        初始化频谱分析器
+        
+        Args:
+            sr: 采样率
+            n_fft: FFT窗口大小
+            hop_length: 跳跃长度
+        """
+        self.sr = sr
+        self.n_fft = n_fft
+        self.hop_length = hop_length
+    
+    def compute_stft(self, audio: np.ndarray) -> np.ndarray:
+        """
+        计算短时傅里叶变换
+        
+        Args:
+            audio: 音频信号
+            
+        Returns:
+            STFT结果
+        """
+        return librosa.stft(audio, n_fft=self.n_fft, hop_length=self.hop_length)
+    
+    def compute_magnitude_spectrum(self, audio: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        计算幅度谱
+        
+        Args:
+            audio: 音频信号
+            
+        Returns:
+            频率轴, 幅度谱
+        """
+        spectrum = fft(audio)
+        magnitude = np.abs(spectrum)
+        freqs = fftfreq(len(audio), 1/self.sr)
+        
+        # 只返回正频率部分
+        positive_freq_idx = freqs >= 0
+        return freqs[positive_freq_idx], magnitude[positive_freq_idx]
+    
+    def spectral_centroid(self, audio: np.ndarray) -> np.ndarray:
+        """
+        计算谱重心
+        
+        Args:
+            audio: 音频信号
+            
+        Returns:
+            谱重心数组
+        """
+        return librosa.feature.spectral_centroid(
+            y=audio, sr=self.sr, hop_length=self.hop_length
+        )[0]
+    
+    def spectral_rolloff(self, audio: np.ndarray, roll_percent: float = 0.85) -> np.ndarray:
+        """
+        计算谱滚降
+        
+        Args:
+            audio: 音频信号
+            roll_percent: 滚降百分比
+            
+        Returns:
+            谱滚降数组
+        """
+        return librosa.feature.spectral_rolloff(
+            y=audio, sr=self.sr, hop_length=self.hop_length, roll_percent=roll_percent
+        )[0]
+    
+    def spectral_flatness(self, audio: np.ndarray) -> np.ndarray:
+        """
+        计算谱平坦度
+        
+        Args:
+            audio: 音频信号
+            
+        Returns:
+            谱平坦度数组
+        """
+        return librosa.feature.spectral_flatness(
+            y=audio, hop_length=self.hop_length
+        )[0]
+    
+    def spectral_contrast(self, audio: np.ndarray, n_bands: int = 6) -> np.ndarray:
+        """
+        计算谱对比度
+        
+        Args:
+            audio: 音频信号
+            n_bands: 频段数量
+            
+        Returns:
+            谱对比度矩阵
+        """
+        return librosa.feature.spectral_contrast(
+            y=audio, sr=self.sr, hop_length=self.hop_length, n_bands=n_bands
+        )
+    
+    def mfcc_features(self, audio: np.ndarray, n_mfcc: int = 13) -> np.ndarray:
+        """
+        提取MFCC特征
+        
+        Args:
+            audio: 音频信号
+            n_mfcc: MFCC系数数量
+            
+        Returns:
+            MFCC特征矩阵
+        """
+        return librosa.feature.mfcc(
+            y=audio, sr=self.sr, n_mfcc=n_mfcc, hop_length=self.hop_length
+        )
+    
+    def mel_spectrogram(self, audio: np.ndarray, n_mels: int = 128) -> np.ndarray:
+        """
+        计算梅尔频谱图
+        
+        Args:
+            audio: 音频信号
+            n_mels: 梅尔滤波器组数量
+            
+        Returns:
+            梅尔频谱图
+        """
+        return librosa.feature.melspectrogram(
+            y=audio, sr=self.sr, n_mels=n_mels, hop_length=self.hop_length
+        )
+    
+    def chroma_features(self, audio: np.ndarray) -> np.ndarray:
+        """
+        提取色度特征
+        
+        Args:
+            audio: 音频信号
+            
+        Returns:
+            色度特征矩阵
+        """
+        return librosa.feature.chroma_stft(
+            y=audio, sr=self.sr, hop_length=self.hop_length
+        )
+
+
+def compute_spectral_features(audio: np.ndarray, sr: int = 22050) -> dict:
+    """
+    计算完整的频域特征集合
+    
+    Args:
+        audio: 音频信号
+        sr: 采样率
+        
+    Returns:
+        包含各种频域特征的字典
+    """
+    analyzer = SpectralAnalyzer(sr=sr)
+    
+    features = {
+        'spectral_centroid': analyzer.spectral_centroid(audio),
+        'spectral_rolloff': analyzer.spectral_rolloff(audio),
+        'spectral_flatness': analyzer.spectral_flatness(audio),
+        'spectral_contrast': analyzer.spectral_contrast(audio),
+        'mfcc': analyzer.mfcc_features(audio),
+        'mel_spectrogram': analyzer.mel_spectrogram(audio),
+        'chroma': analyzer.chroma_features(audio)
+    }
+    
+    return features
+
+
+def frequency_domain_stats(audio: np.ndarray, sr: int = 22050) -> dict:
+    """
+    计算频域统计信息
+    
+    Args:
+        audio: 音频信号
+        sr: 采样率
+        
+    Returns:
+        频域统计信息字典
+    """
+    analyzer = SpectralAnalyzer(sr=sr)
+    freqs, magnitude = analyzer.compute_magnitude_spectrum(audio)
+    
+    # 计算功率谱密度
+    power = magnitude ** 2
+    
+    # 计算统计量
+    stats = {
+        'mean_frequency': np.average(freqs, weights=power),
+        'std_frequency': np.sqrt(np.average((freqs - np.average(freqs, weights=power))**2, weights=power)),
+        'peak_frequency': freqs[np.argmax(magnitude)],
+        'bandwidth': freqs[np.where(power > 0.5 * np.max(power))][-1] - freqs[np.where(power > 0.5 * np.max(power))][0],
+        'spectral_energy': np.sum(power),
+        'spectral_entropy': -np.sum((power/np.sum(power)) * np.log2(power/np.sum(power) + 1e-10))
+    }
+    
+    return stats