load_feature.py

import os.path
import numpy as np
import librosa
import matplotlib.pyplot as plt
from madmom.audio.signal import *

def featureExtract(FILE_NAME):

    try:
        y=  Signal(FILE_NAME, sample_rate=16000,dtype=np.float32,num_channels=1)
        sr = y.sample_rate
        mel_S = librosa.feature.melspectrogram(y, sr=sr, n_fft=1024, hop_length=160, n_mels=80)
        log_mel_S = librosa.power_to_db(mel_S,ref=np.max)
        log_mel_S = log_mel_S.astype(np.float32)
        return log_mel_S

    except Exception as ex:
        print('ERROR: ', ex)
        
def makingTensor(feature,stride):
    num_frames = feature.shape[1]
    x_data = np.zeros(shape=(num_frames, 75, 80, 1))
    total_num = 0
    HALF_WIN_LEN = 75  // 2

    for j in range(HALF_WIN_LEN, num_frames - HALF_WIN_LEN - 2, stride):
        mf_spec = feature[:, range(j - HALF_WIN_LEN, j + HALF_WIN_LEN + 1)]
        x_data[total_num, :, :, 0] = mf_spec.T
        total_num = total_num + 1

    x_data = x_data[:total_num]

    x_train_mean = np.load('./x_data_mean_svad_75.npy')
    x_train_std = np.load('./x_data_std_svad_75.npy')
    x_test = (x_data - x_train_mean) / (x_train_std + 0.0001)

    return x_test


if __name__ == '__main__':
    featureExtract()