开源语音识别软件HTK,对Julius和 Kaldi等系统了解者更佳
一.Installation By Source tarball
Installation process is simple and contains the following points:
1. Download the newest source tarbal from Julius official site – link
2. Unpack the archive for example to your user home directory
3. Configure and install julius by following commands:
<span style="color: rgb(122, 8, 116);"><strong>cd</strong></span> ~<span style="font-weight: bold;">/</span>julius-4.2.2<span style="font-weight: bold;">/</span> <span style="color: rgb(194, 12, 185);"><strong>sudo</strong></span> .<span style="font-weight: bold;">/</span>configure <span style="color: rgb(194, 12, 185);"><strong>sudo</strong></span> <span style="color: rgb(194, 12, 185);"><strong>make</strong></span> <span style="color: rgb(194, 12, 185);"><strong>sudo</strong></span> <span style="color: rgb(194, 12, 185);"><strong>make</strong></span> <span style="color: rgb(194, 12, 185);"><strong>install</strong></span> |
4. Try it by typing in command line:
julius-4.2.1 <span style="font-weight: bold;">//</span>output Julius rev.4.2.2 - based on JuliusLib rev.4.2.2 <span style="color: rgb(122, 8, 116);"><strong>(</strong></span>fast<span style="color: rgb(122, 8, 116);"><strong>)</strong></span> built <span style="font-weight: bold;">for</span> i686-pc-linux Copyright <span style="color: rgb(122, 8, 116);"><strong>(</strong></span>c<span style="color: rgb(122, 8, 116);"><strong>)</strong></span> 1991-2012 Kawahara Lab., Kyoto University Copyright <span style="color: rgb(122, 8, 116);"><strong>(</strong></span>c<span style="color: rgb(122, 8, 116);"><strong>)</strong></span> 1997-2000 Information-technology Promotion Agency, Japan Copyright <span style="color: rgb(122, 8, 116);"><strong>(</strong></span>c<span style="color: rgb(122, 8, 116);"><strong>)</strong></span> 2000-2005 Shikano Lab., Nara Institute of Science and Technology Copyright <span style="color: rgb(122, 8, 116);"><strong>(</strong></span>c<span style="color: rgb(122, 8, 116);"><strong>)</strong></span> 2005-2012 Julius project team, Nagoya Institute of Technology Try <span style="color: rgb(255, 0, 0);">'-setting'</span> <span style="font-weight: bold;">for</span> built-in engine configuration. Try <span style="color: rgb(255, 0, 0);">'-help'</span> <span style="font-weight: bold;">for</span> run <span style="font-weight: bold;">time</span> options. |
5. Last additional thing needed to run Julius smoothly is a julius-voxforge package which can be installed via apt-get by typing in command line:
<span style="color: rgb(194, 12, 185);"><strong>sudo</strong></span> <span style="color: rgb(194, 12, 185);"><strong>apt-get install</strong></span> julius-voxforge |
That’s all now you can start to configure it, because it’s not coming configured just out of the box. I’ll write a post in next few days about basic configuration to which link will be added here.
注意事项:
1.readme当中的命令已经发生了变化:
sudo mkdfa.pl sample
export TMP=/tmp
使用Julius搭建一个语音识别引擎这里主要就是听写程序,可以进行一段语音的连续识别,而且主要是针对中文:
一.语音识别引擎的基本结构
基本所有的开源语音识别引擎都是如下的结构,包括:Sphinx,Julius等,下面主要以Julius为例,说明相关的内容:
1.Acoustic Model(声学模型):用于音素的识别。
所使用的技术:
1) HMM(Hiden Mokov Model):
2) GMM(Gaussian Mixture Model):
3)DFA&NFA:
构建自己的声学模型:
可以使用的工具:
2.Phoneme Dictionary(读音词典):用于词语的识别。
构建自己的读音词典:
可以使用的工具:
3.Language Model(语言模型):用于语句的识别。
构建自己的读音词典:
可以使用的工具:
上面的三者3=>2=>1 依次成依赖关系,只要把上面的三者配置好,一个语音识别引擎就搭建好了,其他的就是利用开发包提供API进行客制化开发了。
二.具体操作
以Julius-Voxforge当中的Sample为例,使用Julius主要就是涉及三个文件:
1.sample.grammar:语言文法。
2. sample.voca:读音词典。
3. julian.jconf:语音识别引擎的主要配置文件。在这里指定相面的三大要素.
至于其他文件:
sample.dfa
sample.dict
sample.term
都是通过命令自动生成的---mkdfa.pl sample。