基于BIC的语音识别模型压缩算法

Speech Recognition Model Compression Algorithm Based on Bayesian Information Criterion

当对HMM(Hidden Markov Model,隐马尔科夫模型)语音模型进行GMM(Gaussian Mixture Model,混合高斯模型)区分训练增加组件时,语音模型的识别率会随着GMM的组件增多而增加,模型的大小也会增加,这就造成了语音模型的臃肿。而在移动端使用本地语音模型进行识别时,存放一个几百兆的模型很不合适。针对上述问题,本文提出将一个GMM组件数较多的语音模型利用BIC准则压缩到指定的组件数,从而在模型大小合适的情况下尽量保证模型的识别率。实验结果表明,使用本方法进行压缩之后的语音识别率比未压缩的相同组件数的语音识别模型的识别率要高。

Recognition rate of speech model will increase with the increase in the number of GMM components, the size of model will increase as well, when making the GMM recognition training for HMM speech model, and it causes model bloated. However, it is unfit for mobile devices while using speech model for recognition to keep greater than hundreds of megabytes in mobile. For this problem, a method for compress speech model based on BIC is presented. This method tries to keep recognition rate of speech model in appropriate to the size of model. Experiments demonstrate that it＇ s applicable and available to achieve the final speech model specified size even ensure recognition rate of speech model as much as possible.