PNCC的python代码

PNCC是一种语音特征提取方法，其Python代码如下：

import numpy as np
from scipy.fftpack import dct

def pncc(signal, samplerate=16000, pre_emphasis=0.97, frame_size=0.025, frame_stride=0.01, N=512, M=512, L=512, alpha=0.5, R=0.95):
    emphasized_signal = np.append(signal[0], signal[1:] - pre_emphasis * signal[:-1])
    frames = np.array([emphasized_signal[i:i + int(frame_size * samplerate)] for i in range(0, len(emphasized_signal) - int(frame_size * samplerate), int(frame_stride * samplerate))])
    frames *= np.hamming(int(frame_size * samplerate))

    NFFT = int(pow(2, np.ceil(np.log2(N))))
    magnitude_spectrum = np.absolute(np.fft.rfft(frames, NFFT))
    power_spectrum = ((1.0 / NFFT) * ((magnitude_spectrum) ** 2))

    filter_bank = np.zeros((M, int(NFFT / 2 + 1)))
    for m in range(1, M + 1):
        f_m_minus = int(np.floor((m - 1) * N / M))
        f_m = int(np.floor(m * N / M))
        f_m_plus = int(np.floor((m + 1) * N / M))

        for k in range(f_m_minus, f_m):
            filter_bank[m - 1, k] = (k - f_m_minus) / (f_m - f_m_minus)
        for k in range(f_m, f_m_plus):
            filter_bank[m - 1, k] = (f_m_plus - k) / (f_m_plus - f_m)

    filter_banks = np.dot(power_spectrum, filter_bank.T)
    filter_banks = np.where(filter_banks == 0, np.finfo(float).eps, filter_banks)
    filter_banks = 20 * np.log10(filter_banks)

    cepstral_coefficents = dct(filter_banks, type=2, axis=1, norm='ortho')[:, :L]
    cepstral_coefficents -= (np.mean(cepstral_coefficents, axis=0) + 1e-8)

    delta_cepstral_coefficents = librosa.feature.delta(cepstral_coefficents)
    delta_delta_cepstral_coefficents = librosa.feature.delta(cepstral_coefficents, order=2)

    numerator = np.power(delta_cepstral_coefficents, 2) + alpha * np.power(delta_delta_cepstral_coefficents, 2)
    denominator = np.ones_like(numerator) + R * np.power(cepstral_coefficents, 2)
    pncc_features = np.divide(numerator, denominator)

    return pncc_features