ニューラルネットの音の正規化処理って何が正解??

import matplotlib.pyplot as plt
plt.figure(figsize=(10, 4))
plt.subplot(2, 1, 1)
plt.plot(wave)
plt.ylabel("Power", size=16)
plt.xlabel("Time", size=16)
plt.title("Original", size=16)
norm_wave = (wave -wave.mean()) / wave.std()
plt.subplot(2, 1, 2)
plt.plot(norm_wave)
plt.ylabel("Power", size=16)
plt.xlabel("Time", size=16)
plt.title("Wave Normalize", size=16)
plt.tight_layout()
plt.savefig("wave.png")

これは、単純に正規化しただけです。生音では分散が小さかったので、正規化により

Powerのオーダーが100倍になったことがわかります。

スペクトログラムに変換

import numpy as np
import librosa

def logmelfilterbank(
    audio,
    sampling_rate,
    fft_size=1024,
    hop_size=256,
    win_length=None,
    window="hann",
    num_mels=80,
    fmin=None,
    fmax=None,
    eps=1e-10,
):
    """Compute log-Mel filterbank feature.

    Args:
        audio (ndarray): Audio signal (T,).
        sampling_rate (int): Sampling rate.
        fft_size (int): FFT size.
        hop_size (int): Hop size.
        win_length (int): Window length. If set to None, it will be the same as fft_size.
        window (str): Window function type.
        num_mels (int): Number of mel basis.
        fmin (int): Minimum frequency in mel basis calculation.
        fmax (int): Maximum frequency in mel basis calculation.
        eps (float): Epsilon value to avoid inf in log calculation.

    Returns:
        ndarray: Log Mel filterbank feature (#frames, num_mels).

    """
    # get amplitude spectrogram
    x_stft = librosa.stft(
        audio,
        n_fft=fft_size,
        hop_length=hop_size,
        win_length=win_length,
        window=window,
        pad_mode="reflect",
    )
    spc = np.abs(x_stft).T  # (#frames, #bins)

    # get mel basis
    fmin = 0 if fmin is None else fmin
    fmax = sampling_rate / 2 if fmax is None else fmax
    mel_basis = librosa.filters.mel(sampling_rate, fft_size, num_mels, fmin, fmax)

    return np.log10(np.maximum(eps, np.dot(spc, mel_basis.T)))


mel = logmelfilterbank(
            wave,
            sampling_rate=16000,
            hop_size=256,
            fft_size=1024,
            window="hann",
            num_mels=128,
            fmin=50,
            fmax=8000,
        )
mel_norm = logmelfilterbank(
            norm_wave,
            sampling_rate=16000,
            hop_size=256,
            fft_size=1024,
            window="hann",
            num_mels=128,
            fmin=50,
            fmax=8000,
        )
mel_norm_norm = (mel - mel.mean(axis=0)) / mel.std(axis=0)

結果と分布を表示

plt.figure(figsize=(16, 8))
plt.subplot(2, 3, 1)
plt.imshow(mel.T, aspect="auto")
plt.xlabel("time", size=16)
plt.ylabel("frequency", size=16)
plt.title("Original", size=16)
plt.subplot(2, 3, 2)
plt.imshow(mel_norm.T, aspect="auto")
plt.xlabel("time", size=16)
plt.ylabel("frequency", size=16)
plt.title("Wave Normalize", size=16)
plt.subplot(2, 3, 3)
plt.imshow(mel_norm_norm.T, aspect="auto")
plt.xlabel("time", size=16)
plt.ylabel("frequency", size=16)
plt.title("LogMel Normalize", size=16)
plt.colorbar()
plt.subplot(2, 3, 4)
_=plt.hist(mel.flatten(), bins=50)
plt.title("Original", size=16)
plt.subplot(2, 3, 5)
_=plt.hist(mel_norm.flatten(), bins=50)
plt.title("Wave Normalize", size=16)
plt.subplot(2, 3, 6)
_=plt.hist(mel_norm_norm.flatten(), bins=50)
plt.title("LogMel Normalize", size=16)
plt.tight_layout()
plt.savefig("hist.png")