s****t 发帖数: 1 | 1 我想着跟你选择的训练单元有关系吧,比如基于phone的,基于sylla
ble的,还是基于word的。
这都快忘了。不过可以确定的是,语言模型与识别器是两个层面的东
西,如果是只是语音数据的识别,当然不用语言模型。如果是要做一
个好的识别器,语言模型少不了,它能够帮你在总多的candidates中
搜索到你最希望的结果。 | e******n 发帖数: 13 | 2
I just built a phone recognizer by TIMIT data. First you
need to decide to use mono-phone or tri-phone as the basic
recognition unit. Then train them as usual. In this step, no
different.
WHen do testing (recognition), you can make decision to use
any language model. For example, you can use the phone-level
bigram as the language model. Different with normal
dictation ASR. The basic unit for LM statistic is phone
instead of word. you need look those phones as the words. |
|