def feature_wav(wav_file, pre_emphasis=0.97, n_filter=40, frame_len_s=0.032, frame_shift_s=0.01):
import numpy as np
from scipy.io import wavfile
from scipy.fftpack import dct
import matplotlib.pyplot as plt
#读取语音数据
fs, sig = wavfile.read(wav_file)
#对语音能量归一化
Ena = sum(sig ** 2)
sig = sig / np.sqrt(Ena)
#预加重
sig = np.append(sig[0], sig[1:] - pre_emphasis * sig[:-1])
#分帧加窗
def framing(frame_len_s, frame_shift_s, fs, sig):
sig_n = len(sig)
frame_len_n, frame_shift_n = int(
round(fs * frame_len_s)), int(round(fs * frame_shift_s))
num_frame = int(
np.ceil(float(sig_n - frame_len_n) / frame_shift_n) + 1)
pad_num = frame_shift_n * (num_frame - 1) + \
frame_len_n - sig_n
pad_zero = np.zeros(int(pad_num)) # ????0
pad_sig = np.append(sig, pad_zero)
frame_inner_index = np.arange(0, frame_len_n)
frame_index = np.arange(0, num_frame) * frame_shift_n
frame_inner_index_extend = np.tile(frame_inner_index, (num_frame, 1))
frame_index_extend = np.expand_dims(frame_index, 1)
each_frame_index = frame_inner_index_extend + frame_index_extend
each_frame_index = each_frame_index.astype(np.int, copy=False)
frame_sig = pad_sig[each_frame_index]
return frame_sig
# frame_len_s = 0.025
# frame_shift_s = 0.01
frame_sig = framing(frame_len_s, frame_shift_s, fs, sig)
window = np.hamming(int(round(frame_len_s * fs)))
frame_sig *= window
#快速傅里叶变换
def stft(frame_sig, nfft=512):
frame_spec = np.fft.rfft(frame_sig, nfft)
frame_mag = np.abs(frame_spec)
frame_pow = (frame_mag ** 2) * 1.0 / nfft
return frame_pow
nfft = 512
frame_pow = stft(frame_sig, nfft)
#梅尔滤波 0全部替换成无穷小 结果以10为底取对数，并乘20
def mel_filter(frame_pow, fs, n_filter, nfft):
mel_min = 0
mel_max = 2595 * np.log10(1 + fs / 2.0 / 700)
mel_points = np.linspace(mel_min, mel_max, n_filter + 2)
hz_points = 700 * (10 ** (mel_points / 2595.0) - 1)
filter_edge = np.floor(hz_points * (nfft + 1) / fs)
fbank = np.zeros((n_filter, int(nfft / 2 + 1)))
for m in range(1, 1 + n_filter):
f_left = int(filter_edge[m - 1])
f_center = int(filter_edge[m])
f_right = int(filter_edge[m + 1])
for k in range(f_left, f_center):
fbank[m - 1, k] = (k - f_left) / (f_center - f_left)
for k in range(f_center, f_right):
fbank[m - 1, k] = (f_right - k) / (f_right - f_center)
filter_banks = np.dot(frame_pow, fbank.T)
filter_banks = np.where(
filter_banks == 0, np.finfo(float).eps, filter_banks)
filter_banks = 20 * np.log10(filter_banks) # dB
return filter_banks
filter_banks = mel_filter(frame_pow, fs, n_filter, nfft)
return filter_banks
在特征提取部分将ta.compliance.kaldi.fbank改为自定义的特征提取函数
由于输出结果是numpy结构的 需要进行转换变为FloatTensor形式
设置了一个输出标志 ，如果这个函数被调用了则输出：DIYFbank
同样在标准化这里也设置个标志：norm
spec-augment:频率掩蔽＋时间掩蔽，忽略时间扭曲
设置标签：mask
训练过程：
三部分都会调用
测试阶段：
只会用到自定义的Fbank和normalization两个部分
词错率为8.401%，与8.276%相比变化增加了0.125%，用自定义Fbank是可行的！
