深度学习vad人声检测之特征生成

得到加噪合成数据后，下面就可以进行特征生成，本博文这里采用的是26维的mfcc特征，当然你也可以根据自己的需求选择不同的特征。准备好了数据，导入直接生成即可，注意这里分帧的方式要和标签制作一致，这里给出的程序当然是一致的。另外需要注意的就是数据均衡的问题，程序中也有体现，具体实现如下：

# -*- coding: utf-8 -*-
from python_speech_features import mfcc
import scipy.io.wavfile as wav
import utils
import json
import math
import numpy as np
from audiolib import audiowrite
from random import shuffle
import os



with open('./feature.json','r') as f:
    params = json.load(f)

    print("params = ",params)


def read_label_info(path):
    f = open(path,'r')
    lines = f.readlines()

    label = []
    for line in lines:
        line = line.strip().split(' ')
        tmp = []
        for l in line:
            tmp.append(int(l))
        label.append(tmp)
    
    f.close()

    return label



def time2frame(labels,frame_len = 400,frame_step = 160):
    frame_label = []

    for label in labels:
        tmp = []
        for l in label:
            numframe = 1 + int(math.ceil((1.0 * l - frame_len) / frame_step))
            tmp.append(numframe)
        frame_label.append(tmp)

    return frame_label


def get_frame_labels(frame_indexs,length):

    frame_labels = np.zeros((length,1),dtype = int)
    for ind in frame_indexs:
        frame_labels[ind[0] : ind[1],:] = 1
    
    return frame_labels


def write_frame_labels(frame_labels,fp):

    for label in frame_labels:
        fp.write(str(int(label)) + ' ')
    fp.write('\n')

    return fp


def write_frame_feats(frame_feats,path):

    fp = open(path,'w')
    for feat in frame_feats:
        for f in feat:
            fp.write(str(f) + ' ')
        fp.write('\n')
    
    fp.close()





def get_mfcc_features(params,mode,flag = 0):
    '''
    语音加噪数据生成mfcc特征
    '''
    if flag:
        print('--->>>train feat dirs = ',params['train']['feats_data_dir'])
        utils.del_folder(params['train']['feats_data_dir'])
        utils.del_folder(params['test']['feats_data_dir'])
        utils.del_folder(params['valid']['feats_data_dir'])


    utils.create_folder(params['train']['feats_data_dir'])
    utils.create_folder(params['test']['feats_data_dir'])
    utils.create_folder(params['valid']['feats_data_dir'])

    data_dir = params[mode]['noisy_data_dir']
    #print('data_dir = ',data_dir)
    data_path,names = utils.read_audio_file1(data_dir,'.wav')
    data_len = len(data_path)
    print('data_len = ',data_len)

    #对数据进行划分
    train_len = int(data_len * 0.7)
    valid_len = int(data_len * 0.85)

    train_data_len = train_len
    valid_data_len = valid_len - train_len
    test_data_len = data_len - valid_len

    acc_sum = 0
    pos_sum = 0

    cnt = 0
    ind = [i for i in range(data_len)]
    shuffle(ind)

    for i in ind:
        dp_p = data_path[i]
        name = names[i]
 
        (fs,sig) = wav.read(dp_p)

        mfcc_feat = mfcc(sig,fs,numcep = params[mode]['mel_num'])
        lb_len = len(mfcc_feat)

        #print('lb_len = ',lb_len)
        #print('mfcc_feat.shape = ',mfcc_feat.shape)

        print('cnt = ',cnt)

        #将生成的特征和标签信息写入文本保存
        if cnt < train_len:
            save_feat_path = params['train']['feats_data_dir'] + '/' + name + '.npy'
            np.save(save_feat_path,mfcc_feat)

            #write_frame_feats(mfcc_feat,save_feat_path)

        elif cnt > train_len and cnt < valid_len:
            save_feat_path = params['valid']['feats_data_dir'] + '/' + name + '.npy'
            np.save(save_feat_path,mfcc_feat)
            #write_frame_feats(mfcc_feat,save_feat_path)

        else:
            save_feat_path = params['test']['feats_data_dir'] + '/' + name + '.npy'
            np.save(save_feat_path,mfcc_feat)
            #write_frame_feats(mfcc_feat,save_feat_path)

        
        cnt = cnt + 1
        
    return train_data_len,valid_data_len,test_data_len


class CombineNoise():
    '''
    噪声数据拼接
    '''

    def __init__(self,params,sample_length):

        self.noise_data_dir = params['noise']['noise_data_dir']
        self.sample_length = sample_length
        self.fs = params['noise']['sample_rate']

        self.data_paths,self.names = utils.read_audio_file1(self.noise_data_dir,'.wav')
        

    def combine_noise(self):

        remain_length = self.sample_length * self.fs
        sig_combine = np.zeros((remain_length),dtype = np.int16)

        data_len = len(self.data_paths) - 1
        ind = [i for i in range(data_len)]
        shuffle(ind)

        cnt = 0
        idx = 0

        while remain_length > 0:
            index = ind[idx]
            (fs,sig) = wav.read(self.data_paths[index])
            sig_len = len(sig)
            if sig_len > remain_length:
                sig_combine[:(cnt + remain_length)] = np.concatenate([sig_combine[:cnt],sig[:remain_length]])
                cnt = cnt + remain_length
            else:
                sig_combine[:(cnt + sig_len)] = np.concatenate([sig_combine[:cnt],sig])
                cnt = cnt + sig_len
            
            #index = index + 1
            idx = idx + 1
            remain_length = remain_length - sig_len
        
        return sig_combine,self.fs


    def get_combine_noise(self):

        index = 0
        data_len = len(self.data_paths)
        cnt = 0

        while index < data_len:
            sig_combine,_ = self.combine_noise(index)
            save_path = self.combine_data_dir + '/' + str(cnt) + '.wav'
            audiowrite(save_path,sig_combine,self.fs)
            cnt = cnt + 1


def get_noise_mfcc_features0(params,mode,length,fp,flag = 0):
    '''
    噪声数据生成mfcc特征
    '''
    data_dir = params['noise']['noise_data_dir']
    data_path,names = utils.read_audio_file1(data_dir,'.wav')

    data_len = len(data_path)
    print('data_len = ',data_len)
    ind = [i for i in range(data_len)]
    shuffle(ind)
    #ind = [n for n in range(16000)]

    comNoise = CombineNoise(params,params['noise']['wav_time'])

    cnt = 0
    for j in range(length):

        print('cnt = ',cnt)

        sig,fs = comNoise.combine_noise()

        mfcc_feat = mfcc(sig,fs,numcep = params[mode]['mel_num'])
        lb_len = len(mfcc_feat)
        print('mfcc_feat.shape = ',mfcc_feat.shape)

        #每帧的标签为0
        frame_labels = np.zeros((lb_len,1),dtype = int)

        #写入每个音频文件对应帧特征和对应帧标签
        save_feat_path = params[mode]['feats_data_dir'] + '/' + mode + str(cnt) + '_noisy_snr100.npy'
        np.save(save_feat_path,mfcc_feat)
        #write_frame_feats(mfcc_feat,save_feat_path)

        label_name = mode + str(cnt) + ':'
        fp.write(label_name)
        fp = write_frame_labels(frame_labels,fp)

        cnt = cnt + 1

    return fp



def get_noise_mfcc_features1(params,mode,length,fp,idx = 0,flag = 0):
    '''
    噪声数据生成mfcc特征
    '''
    data_dir = params['noise']['noise_data_dir']
    data_path,names = utils.read_audio_file1(data_dir,'.wav')

    data_len = len(data_path)
    print('data_len = ',data_len)
    ind = [i for i in range(data_len)]
    shuffle(ind)
    #ind = [n for n in range(16000)]

    #comNoise = CombineNoise(params,params['noise']['wav_time'])

    sub_data_path = data_path[idx:(idx + length)]

    cnt = 0
    for dpath in sub_data_path:

        print('cnt = ',cnt)

        (fs,sig) = wav.read(dpath)
        #sig,fs = comNoise.combine_noise()

        mfcc_feat = mfcc(sig,fs,numcep = params[mode]['mel_num'])
        lb_len = len(mfcc_feat)
        #print('mfcc_feat.shape = ',mfcc_feat.shape)

        #每帧的标签为0
        frame_labels = np.zeros((lb_len,1),dtype = int)

        #写入每个音频文件对应帧特征和对应帧标签
        save_feat_path = params[mode]['feats_data_dir'] + '/' + mode + str(cnt) + '_noisy_snr100.npy'
        np.save(save_feat_path,mfcc_feat)

        label_name = mode + str(cnt) + ':'
        fp.write(label_name)
        fp = write_frame_labels(frame_labels,fp)

        cnt = cnt + 1

    return fp





import shutil
def copy_files1(sources_path,dst_path1,dst_path2,dst_path3,fmt = '.wav'):

    files,names = utils.read_audio_file1(sources_path,fmt)
    sources_label_path = sources_path.replace('feats','labels')
    labels,lnames = utils.read_audio_file1(sources_label_path,fmt)

    len1 = 25097
    len2 = len1 + 5904
    len3 = len2 + 5906



    cnt = 0
    for (file,label) in zip(files,labels):
        if cnt < len1:
            shutil.copy(file,dst_path1)
            dst_label_path1 = dst_path1.replace('feats','labels')
            shutil.copy(label,dst_label_path1)
        elif cnt > len1 and cnt < len2:
            shutil.copy(file,dst_path2)
            dst_label_path2 = dst_path2.replace('feats','labels')
            shutil.copy(label,dst_label_path2)
        elif cnt > len2 and cnt < len3:
            shutil.copy(file,dst_path3)
            dst_label_path3 = dst_path3.replace('feats','labels')
            shutil.copy(label,dst_label_path3)
        
        cnt = cnt + 1


def copy_files2(sources_path,dst_path,fmt = '.wav'):

    files,names = utils.read_audio_file1(sources_path,fmt)
    sources_label_path = sources_path.replace('feats','labels')
    labels,lnames = utils.read_audio_file1(sources_label_path,fmt)

    for (fi,label) in zip(files,labels):
        shutil.copy(fi,dst_path)
        dst_label_path = dst_path.replace('feats','labels')
        shutil.copy(label,dst_label_path)
    


if __name__ == '__main__':

    print('---------get pos features-------')
    ratio = 0.78   #ratio可以通过vad标签制作过程统计获得
    train_data_len,valid_data_len,test_data_len = get_mfcc_features(params,'noisy',0)
    
    print('---------add neg features-------')
    '''
    train_data_len = 12600
    valid_data_len = 2699
    test_data_len = 2702
    '''
    train_neg_add = int(train_data_len * ratio - train_data_len * (1 - ratio))
    valid_neg_add = int(valid_data_len * ratio - valid_data_len * (1 - ratio))
    test_neg_add = int(test_data_len * ratio - test_data_len * (1 - ratio))
    
    # fp = open(params['labels']['noise_label_dir'],'w')
    # fp = get_noise_mfcc_features0(params,'train',train_neg_add,fp)
    # fp = get_noise_mfcc_features0(params,'valid',valid_neg_add,fp)
    # fp = get_noise_mfcc_features0(params,'test',test_neg_add,fp)


    fp = open(params['labels']['noise_label_dir'],'w')
    ind = 0
    fp = get_noise_mfcc_features1(params,'train',train_neg_add,fp,ind)
    ind = ind + train_neg_add
    fp = get_noise_mfcc_features1(params,'valid',valid_neg_add,fp,ind)
    ind = ind + valid_neg_add
    fp = get_noise_mfcc_features1(params,'test',test_neg_add,fp,ind)

至此，特征生成就完成了，水平有限，不当之处请指教，谢谢！