梅尔频谱(mel-spectrogram)提取,griffin_lim声码器【python代码分析】

爱被打了一巴掌 2022-12-28 12:59 152阅读 0赞

在语音分析，合成，转换中，第一步往往是提取语音特征参数。  
利用机器学习方法进行上述语音任务，常用到梅尔频谱。  
本文介绍从音频文件提取梅尔频谱，和从梅尔频谱变成音频波形。

**从音频波形提取Mel频谱：**

1.  对音频信号预加重、分帧和加窗
2.  对每帧信号进行短时傅立叶变换STFT，得到短时幅度谱
3.  短时幅度谱通过Mel滤波器组得到Mel频谱

**从Mel频谱重建音频波形**

1.  Mel频谱转换成幅度谱
2.  griffin\_lim声码器算法重建波形
3.  去加重

> 声码器有很多种，比如world,straight等，但是griffin\_lim是特殊的，它不需要相位信息就可以重频谱重建波形，实际上它根据帧之间的关系估计相位信息。和成的音频质量也较高，代码也比较简单。

### 音频波形 到 mel-spectrogram ###

sr = 24000 # Sample rate.
    n_fft = 2048 # fft points (samples)
    frame_shift = 0.0125 # seconds
    frame_length = 0.05 # seconds
    hop_length = int(sr*frame_shift) # samples.
    win_length = int(sr*frame_length) # samples.
    n_mels = 512 # Number of Mel banks to generate
    power = 1.2 # Exponent for amplifying the predicted magnitude
    n_iter = 100 # Number of inversion iterations
    preemphasis = .97 # or None
    max_db = 100
    ref_db = 20
    top_db = 15

def get_spectrograms(fpath):
        '''Returns normalized log(melspectrogram) and log(magnitude) from `sound_file`.
        Args:
          sound_file: A string. The full path of a sound file.
    
        Returns:
          mel: A 2d array of shape (T, n_mels) <- Transposed
          mag: A 2d array of shape (T, 1+n_fft/2) <- Transposed
     '''
        # Loading sound file
        y, sr = librosa.load(fpath, sr=sr)
    
        # Trimming
        y, _ = librosa.effects.trim(y, top_db=top_db)
    
        # Preemphasis
        y = np.append(y[0], y[1:] - preemphasis * y[:-1])
        
        # stft
        linear = librosa.stft(y=y,
                              n_fft=n_fft,
                              hop_length=hop_length,
                              win_length=win_length)
    
        # magnitude spectrogram
        mag = np.abs(linear)  # (1+n_fft//2, T)
    
        # mel spectrogram
        mel_basis = librosa.filters.mel(sr, n_fft, n_mels)  # (n_mels, 1+n_fft//2)
        mel = np.dot(mel_basis, mag)  # (n_mels, t)
    
        # to decibel
        mel = 20 * np.log10(np.maximum(1e-5, mel))
        mag = 20 * np.log10(np.maximum(1e-5, mag))
    
        # normalize
        mel = np.clip((mel - ref_db + max_db) / max_db, 1e-8, 1)
        mag = np.clip((mag - ref_db + max_db) / max_db, 1e-8, 1)
    
        # Transpose
        mel = mel.T.astype(np.float32)  # (T, n_mels)
        mag = mag.T.astype(np.float32)  # (T, 1+n_fft//2)
    
        return mel, mag

### mel-spectrogram 到 音频波形 ###

def melspectrogram2wav(mel):
        '''# Generate wave file from spectrogram'''
        # transpose
        mel = mel.T
    
        # de-noramlize
        mel = (np.clip(mel, 0, 1) * max_db) - max_db + ref_db
    
        # to amplitude
        mel = np.power(10.0, mel * 0.05)
        m = _mel_to_linear_matrix(sr, n_fft, n_mels)
        mag = np.dot(m, mel)
    
        # wav reconstruction
        wav = griffin_lim(mag)
    
        # de-preemphasis
        wav = signal.lfilter([1], [1, -preemphasis], wav)
    
        # trim
        wav, _ = librosa.effects.trim(wav)
    
        return wav.astype(np.float32)
    
    def spectrogram2wav(mag):
        '''# Generate wave file from spectrogram'''
        # transpose
        mag = mag.T
    
        # de-noramlize
        mag = (np.clip(mag, 0, 1) * max_db) - max_db + ref_db
    
        # to amplitude
        mag = np.power(10.0, mag * 0.05)
    
        # wav reconstruction
        wav = griffin_lim(mag)
    
        # de-preemphasis
        wav = signal.lfilter([1], [1, -preemphasis], wav)
    
        # trim
        wav, _ = librosa.effects.trim(wav)
    
        return wav.astype(np.float32)

几个辅助函数：

def _mel_to_linear_matrix(sr, n_fft, n_mels):
        m = librosa.filters.mel(sr, n_fft, n_mels)
        m_t = np.transpose(m)
        p = np.matmul(m, m_t)
        d = [1.0 / x if np.abs(x) > 1.0e-8 else x for x in np.sum(p, axis=0)]
        return np.matmul(m_t, np.diag(d))
    
    def griffin_lim(spectrogram):
        '''Applies Griffin-Lim's raw.
        '''
        X_best = copy.deepcopy(spectrogram)
        for i in range(n_iter):
            X_t = invert_spectrogram(X_best)
            est = librosa.stft(X_t, n_fft, hop_length, win_length=win_length)
            phase = est / np.maximum(1e-8, np.abs(est))
            X_best = spectrogram * phase
        X_t = invert_spectrogram(X_best)
        y = np.real(X_t)
    
        return y
    
    
    def invert_spectrogram(spectrogram):
        '''
        spectrogram: [f, t]
        '''
        return librosa.istft(spectrogram, hop_length, win_length=win_length, window="hann")

**预加重：**  
语音信号的平均功率谱受声门激励和口鼻辐射影响，高频端约在800HZ以上按6dB/倍频程衰落，预加重的目的是提升高频成分，使信号频谱平坦化，以便于频谱分析或声道参数分析．