[Audio processing] 数据集生成 & 性别年龄分类训练 Python

1、重命名，Python中文路径各种错误，所以需要先将所有文件的路径名全都改成中文。用的是MAC系统，所以WIN下的命令行批处理没法解决，所以用C来完成

//  Created by Carl on 16.
//  Copyright (c) 2016年 Carl. All rights reserved.
//

#include <iostream>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <dirent.h>
#include <unistd.h>
using namespace std;

void getFileList()
{
    string sourceDir = "/Users/karl/Work/database/rawdata/children_CN/";
    string targetDir = "/Users/karl/Work/database/rawdata/children/";
    DIR *dir;
    struct dirent *ptr;
    int i = 0;
    if ((dir=opendir(sourceDir.c_str())) == NULL)
    {
        perror("Open dir error...");
        exit(1);
    }
    while ((ptr=readdir(dir)) != NULL)
    {
        if(strcmp(ptr->d_name,".")==0 || strcmp(ptr->d_name,"..")==0)    ///current dir OR parrent
            continue;
        else if(ptr->d_type == 8)
        {
            printf("%s %s\n",(sourceDir + ptr->d_name).c_str(),(targetDir + to_string(i) + ".wav").c_str());
            if(rename((sourceDir + ptr->d_name).c_str(), (targetDir + to_string(i++) + ".wav").c_str())<0)
                cout<<"error"<<endl;
            else
                cout<<"ok"<<endl;
        }
        
    }
    return;
}

int main() {
    getFileList();
    return 1;
}

2、然后再使用FFMPEG那篇文章写的Python代码，将所有音频文件转成统一格式

#coding=utf-8
#!/usr/bin/env python
'''CREATED:2016-03-08
Use example of ffmpeg
'''
import argparse
import sys
import os
import string
import subprocess as sp

#Full path of ffmpeg
FFMPEG_BIN = "/Users/karl/Documents/python/audio/tool/ffmpeg"
#Full path of sourceDir
sourceDir = "/Users/karl/Work/database/rawdata/male/"
#Full path of targetDir
targetDir = "/Users/karl/Work/database/age/male/"
#Channel setting 1 for mono
ac = 1
#Sample frequency
sf = 16000
#Extension setting
ext = 'wav'

def convert(sourceDir, targetDir, ac, sf, ext):
    i = 0
    if not os.path.exists(targetDir):
        os.mkdir(targetDir)
    files = os.listdir(sourceDir)
    for f in files:
        if f.endswith('.wav'):
            command = [ FFMPEG_BIN,
                       '-i', os.path.join(sourceDir, f),
                       '-ac', str(ac),
                       '-ar', str(sf), os.path.join(targetDir, str(i) + "." + ext)]
            i += 1
            print command
            pipe = sp.Popen(command, stdout = sp.PIPE, bufsize = 10**8)

if __name__ == '__main__':
    convert(sourceDir, targetDir, ac, sf, ext)

3、用时域上RMS去除静音帧(Optional)

#---Cut the silent head and tail of audio
def rmsdemo(y):
    return np.sqrt((y**2).mean())

def cutheadntail(y, winlen, threshold):
    totallen = y.shape[0]
    num = totallen / winlen
    i = 1
    j = num
    for i in range(num):
        if rmsdemo(y[i * winlen : (i + 1) * winlen - 1]) > threshold:
            break
    for j in range(-1,0,-1):
        if rmsdemo(y[i * winlen : (i + 1) * winlen - 1]) > threshold or j == i:
            break
    #percentage = (j - i + 1) * 1.0 / num;
    #print(i, j, percentage)
    yy = y[i * winlen : (j + 1) * winlen - 1]
    return yy

4、用librosa提取特征，包括MFCC、DMFCC

from __future__ import print_function
import argparse
import sys
import os
import pprint
import sklearn as sl
import numpy as np
import librosa
import librosa.feature.spectral as f
import svmutil

#---Feature extraction and store, including MFCC, DMFCC
def mfcclist(data_dir):
    m = []
    dm = []
    for i in range(300):
        filepath = os.path.join(data_dir, str(i) + '.wav')
        print(filepath)
        am, adm = mfccfile(filepath)
        m.append(am)
        dm.append(adm)
        i += 1
    np.savetxt("TrainFemaleMFCC",m,fmt='%s',newline='\n')
    np.savetxt("TrainFemaleDMFCC",dm,fmt='%s',newline='\n')
    #print(m)
    #print(dm)
'''
    fout = open(output_file,'w')
    fout.write(str(am) + '\n')
    fout.write(str(adm))
    fout.close()
'''

def mfccfile(input_file):
    print('Loading ', input_file)
    y, sr = librosa.load(input_file)
    M = f.mfcc(y, sr, None, 13)
    DM = M[::,1::] - M[::,0:-1:1]
    am = np.mean(M, axis = 1)
    adm = np.mean(DM, axis = 1)
    return (am, adm)

#---Loading stored features file
def loadfeatures(features_file):
    fin = open(features_file, 'r')
    features = [map(float,ln.strip().split(' '))
                for ln in fin.read().splitlines() if ln.strip()]
                #pprint.pprint(features)
    print(features)

5、用libsvm训练和预测，包括归一化

#---SVM training and predicting process
def svmtraindemo(x, modelname, scalar):
    x = scalar.transform(x)
    #x = sl.preprocessing.scale(x)
    x = x.tolist()
    print(x)
    y = [1.0] * 300 + [1] * 300 + [-1.0] * 600
    model = svm_train(y, x, '-b 1')
    svm_save_model(modelname + str(0), model)
    p_label, p_acc, p_val = svm_predict(y[:1200], x[:1200], model, '-b 1')

def svmpredictdemo(x, modelname, scalar):
    x = scalar.transform(x)
    #x = sl.preprocessing.scale(x)
    x = x.tolist()
    print(len(x))
    y = [1.0] * 100 + [1] * 100 + [-1.0] * 200
    m = svm_load_model(modelname + str(0))
    print(p_label)
    p_label, p_acc, p_val = svm_predict(y[:400], x[:400], m, '-b 1')

 附：

1、经过试验，发现用无监督的方式，准确来说是基于规则的方式分辨男、女、小孩的声音还是不太靠谱，频域上的分布还是用有监督的方式自己学习应该更可靠。

2、用有噪音的推无噪音的小孩，准确率80%，无噪音推有噪音的，准确率才60+%，所以训练还是最好用噪音环境的数据集吧，之前想的是训练应该用无噪音的样本还是太天真了。其实混合起来效果还不错。

3、男女的准确率也就80%，样本分布还是比较好，而且均有噪音，估计在实际应用中效果也不会比80%差太远。