基於各種分類算法的說話人識別（年齡段識別）

基於各種分類算法的語音分類（年齡段識別）

概述

實習期間作為幫手打雜進行了一段時間的語音識別研究，內容是基於各種分類算法的語音的年齡段識別，總結一下大致框架，基本思想是：

• 獲取語料庫
  TIMIT

• 提取數據特征，進行處理
  MFCC/i-vector
  LDA/PLDA/PCA

• 語料提取，基於分類算法進行分類
  SVM/SVR/GMM/GBDT...

用到的工具有HTK(C,shell)/Kaldi(C++,shell)/LIBSVM(Python)/scikit-learn(Python)

獲取語料庫

TIMIT語料庫 http://www.cnblogs.com/welen/p/3782804.html

PS：

• TIMIT的語料語音（即子文件夾下的WAV文件）是SPHERE文件，可以用Kaldi轉換
• TIMIT/DOC/SPKRINFO.TXT中為speaker信息，作為分類條件

提取數據特征，進行處理

將SPHERE文件轉換為WAV文件

Kaldi中tools下有SPHERE文件轉換工具sph2pipe.exe

cd kaldi/kaldi-trunk/tools/sph2pipe_v2.5/

轉換方法

sph2pipe -f wav sourcefile targetfile

用re_sph2pipe.py腳本生成sph2pipe轉換文件

#encoding="utf-8"
import os
import os.path
rootdir = "E:/vc/TIMIT"
timitpath = "/home/zhangzd/kaldi/kaldi-trunk/TIMIT"
sph2pipepath = "/home/zhangzd/kaldi/kaldi-trunk/tools/sph2pipe_v2.5/sph2pipe"
f = open('E:/vc/data/mfcc/make_sph2pipe_file.txt','w')
for root,dirs,files in os.walk(rootdir):
 for fn in files:
    if fn[len(fn)-3:len(fn)]=='WAV':
        sourcefile = timitpath+root[len(rootdir):]+"/"+fn
        targetfile = root[len(root)-5:len(root)]+"_"+fn
        s = sph2pipepath + " -f wav " + sourcefile+" "+targetfile+"\n"
        f.write(s.replace('\\','/'))
f.close()

得到的轉換文件make_sph2pipe_file.txt如下

/home/zhangzd/kaldi/kaldi-trunk/tools/sph2pipe_v2.5/sph2pipe -f wav /home/zhangzd/kaldi/kaldi-trunk/TIMIT/CONVERT/SA1.WAV NVERT_SA1.WAV
/home/zhangzd/kaldi/kaldi-trunk/tools/sph2pipe_v2.5/sph2pipe -f wav /home/zhangzd/kaldi/kaldi-trunk/TIMIT/TIMIT/TEST/DR1/FAKS0/SA1.WAV FAKS0_SA1.WAV
/home/zhangzd/kaldi/kaldi-trunk/tools/sph2pipe_v2.5/sph2pipe -f wav /home/zhangzd/kaldi/kaldi-trunk/TIMIT/TIMIT/TEST/DR1/FAKS0/SA2.WAV FAKS0_SA2.WAV
...

最后在linux下執行shell命令

#!bin/sh
while read line
do
  echo $line
done make_sph2pipe_file.txt

PS：
f.write(s.replace('\\','/'))是因為在windows下用\\表示路徑，在linux下用/表示

在Kaldi中生成MFCC特征

解析/home/zhangzd/kaldi/kaldi-trunk/egs/wsj/s5/steps/make_mfcc.sh
中提取特征代碼為

$cmd JOB=1:$nj $logdir/make_mfcc_${name}.JOB.log \
  compute-mfcc-feats --verbose=2 --config=$mfcc_config \
   scp,p:$logdir/wav_${name}.JOB.scp ark:- \| \
    copy-feats --compress=$compress ark:- \
    ark,scp:$mfccdir/raw_mfcc_$name.JOB.ark,$mfccdir/raw_mfcc_$name.JOB.scp \
    || exit 1;

即生成MFCC命令為

compute-mfcc-feats --verbose=2 --config=config.txt scp,p:scp.txt ark:-|copy-feats ark:- ark,scp:mfcc.ark,mfcc.scp

config.txt格式為

--use-energy=false   # only non-default option.
...

scp.txt格式為

FAKS0_SA1 /home/zhangzd/kaldi/kaldi-trunk/src/test/FAKS0_SA1.WAV

mfcc.scp格式為

FAKS0_SA1 /home/zhangzd/kaldi/kaldi-trunk/src/test/mfcc.ark

mfcc.ark會自動生成

HTK中生成MFCC特征

HTK更為簡單

HCopy -c config.txt -S scp.txt

config.txt格式為

SOURCEFORMAT = WAV               # Gives the format of the speech files
TARGETKIND = MFCC_0_D_A       # Identifier of the coefficients to use

# Unit = 0.1 micro-second :
WINDOWSIZE = 250000.0          # = 25 ms = length of a time frame
TARGETRATE = 100000.0          # = 10 ms = frame periodicity

NUMCEPS = 12               # Number of MFCC coeffs (here from c1 to c12)
USEHAMMING = T           # Use of Hamming function for windowing frames
PREEMCOEF = 0.97                # Pre-emphasis coefficient
NUMCHANS = 26                 # Number of filterbank channels
CEPLIFTER = 22                   # Length of cepstral liftering
ENORMALIZE = T

scp.txt格式為

E:\vc\data\timit\FADG0_SA1.WAV E:\vc\data\mfcc\FADG0_SA1.mfcc
E:\vc\data\timit\FADG0_SA2.WAV E:\vc\data\mfcc\FADG0_SA2.mfcc
E:\vc\data\timit\FADG0_SI1279.WAV E:\vc\data\mfcc\FADG0_SI1279.mfcc
...

其他

• i-vector
• vad