ubuntu16.04下用HTK构建语音识别系统

假设已经安装好了HTK并设置好了环境变量。使用的数据集是timit数据集。

数据准备（Data Preparation）

建立任务语法（Task Grammar）

从timit中抽取数据

抽取发音字典

使用如下的python程序从timit中抽取发音词典并修改成为HTK所使用的格式。

def timit_dict_to_htk_dict(timit_dict_file,htk_dict_file):
    timit_dicts = [line.strip().split('  ') for line in open(timit_dict_file,'r') if not line.startswith(';')]
    source_dict = {
   }
    for line in timit_dicts:
        word,pronun = line[0].upper().split('~')[0].strip(),line[1]
        for c in ['.']:
            word = word.strip(c)
        if word.startswith("'"):
            word = word[1:]
        if word not in source_dict:
            source_dict[word] = []
        source_dict[word].append(pronun)
    
    new_word_pronun = []
    for word in source_dict:
        for pronun in source_dict[word]:
            new_word_pronun.append([word,pronun])
    new_word_pronun = sorted(new_word_pronun,key= lambda x:x[0])
        
        
    htk_dicts = [[line[0].upper(),line[1].replace('/','')] for line in new_word_pronun]
    htk_dicts = sorted(htk_dicts,key = lambda x:x[0])
    outf = open(htk_dict_file,'w')
    outf.write('\n'.join([' '.join(line) for line in htk_dicts]) + '\n')
    outf.close()


timit_path = '/home/wd/D/Research/dataset/timit'
source_dict_file = '%s/timit/doc/timitdic.txt' % timit_path
htk_dict_file = 'data/htkdic'
timit_dict_to_htk_dict(source_dict_file, htk_dict_file)

程序会创建文件data/htkdic，其内容为：

-KNACKS n ae1 k s
-UPMANSHIP ah1 p m ax n sh ih p
-UPS ah p s
-ZAGGED z ae1 g d
A ax
ABBREVIATE ax b r iy1 v iy ey2 t
ABDOMEN ae1 b d ax m ax n
ABIDES ax b ay1 d z
ABILITY ax b ih1 l ix t iy
ABLE ey1 b el
...

在程序目录中创建data文件夹，使用如下的python脚本来提取timit的一个子集。

import os
Delete_Letter = ['"',':',',','!','?','.','~',';']
def wav_trans(timit_path,data_set,htk_dict_file,prompts_file):
    htk_dict_vocabulary = [line.strip().split(' ')[0] for line in open(htk_dict_file,'r')]
    
    wav_trans_list = []
    path_1 = '%s/timit/%s' % (timit_path,data_set)
    for item in os.listdir(path_1):
        path_2 = '%s/%s' % (path_1,item)
        for sub_item in os.listdir(path_2):
            path_3 = '%s/%s' % (path_2,sub_item)
            for file in os.listdir(path_3):
                file_path = '%s/%s' % (path_3,file)
                if file_path.endswith('.txt'):
                    trans = open(file_path,'r').readline().strip().split(' ')[2:]
                    trans = [i[1:].upper() if i.startswith("'") else i.upper() for i in trans]
                    keep = True
                    trans = ' '.join(trans)
                    for c in Delete_Letter:
                        trans = trans.replace(c,'')
                    trans = trans.split(' ')
                    for item in trans:
                        if item not in htk_dict_vocabulary:
                            keep = False
                    if keep:
                        trans = ' '.join(trans)
                            
                        file_path = file_path[:-3] + 'wav'
                        wav_trans_list.append([file_path,trans])
                
    outf = open(prompts_file,'w')
    outf.write('\n'.join(['\t'.join(line) for line in wav_trans_list][:]))
    outf.close()
    
data_set = 'train'
prompts_file = 'data/%sprompts' % data_set
wav_trans(timit_path, data_set,htk_dict_file,prompts_file)

程序运行结束后，data/中有一个文件trainprompts,其中内容为：

/home/wd/D/Research/dataset/timit/timit/train/dr1/fcjf0/sa1.wav	SHE HAD YOUR DARK SUIT IN GREASY WASH WATER ALL YEAR
/home/wd/D/Research/dataset/timit/timit/train/dr1/fcjf0/sa2.wav	DON'T ASK ME TO CARRY AN OILY RAG LIKE THAT
/home/wd/D/Research/dataset/timit/timit/train/dr1/fcjf0/si1027.wav	EVEN THEN IF SHE TOOK ONE STEP FORWARD HE COULD CATCH HER
/home/wd/D/Research/dataset/timit/timit/train/dr1/fcjf0/si1657.wav	OR BORROW SOME MONEY FROM SOMEONE AND GO HOME BY BUS
/home/wd/D/Research/dataset/timit/timit/train/dr1/fcjf0/si648.wav	A SAILBOAT MAY HAVE A BONE IN HER TEETH ONE MINUTE AND LIE BECALMED THE NEXT
/home/wd/D/Research/dataset/timit/timit/train/dr1/fcjf0/sx127.wav	THE EMPEROR HAD A MEAN TEMPER
/home/wd/D/Research/dataset/timit/timit/train/dr1/fcjf0/sx217.wav	HOW PERMANENT ARE THEIR RECORDS
/home/wd/D/Research/dataset/timit/ti