一种改进的DNN-HMM的语音识别方法

An improved speech recognition method based on DNN-HMM model

针对深度神经网络与隐马尔可夫模型(DNN-HMM)结合的声学模型在语音识别过程中建模能力有限等问题,提出了一种改进的DNN-HMM模型语音识别算法.首先根据深度置信网络(DBN)结合深度玻尔兹曼机(DBM),建立深度神经网络声学模型,然后提取梅尔频率倒谱系数(MFCC)和对数域的Mel滤波器组系数(Fbank)作为声学特征参数,通过TIMIT语音数据集进行实验.实验结果表明:结合了DBM的DNN-HMM模型相比DNN-HMM模型更具优势,其中,使用MFCC声学特征在词错误率与句错误率方面分别下降了1.26% 和0.20%.此外,使用默认滤波器组的Fbank特征在词错误率与句错误率方面分别下降了0.48% 和0.82%,并且适量增加滤波器组可以降低错误率.总之,研究取得句错误率与词错误率分别降低到21.06% 和3.12% 的好成绩.

The acoustic model combined with deep neural network and hidden Markov model (DNN-HMM) has been used extensively in today's speech recognition system.In this paper, an improved DNN-HMM model speech recognition algorithm is proposed. First, a deep neural network acoustic model is built by the deep belief network (DNN) and the deep Boltzmann machine (DBM). Then the Mel frequency cepstral coefficient (MFCC) and the log filter coefficient of the log domain (Fbank) are extracted as an acoustic feature parameter. Finally, the experiment is performed on the TIMIT speech data set. The experimental results show that the DNN-HMM model combined with DBM has more advantages than DNN-HMM model, in which the MFCC acoustic features can reduce the word error rate and sentence error rate by 1.26% and 0.20% respectively. Moreover, using the Fbank feature default filter group rate decreased the word error rate and sentence error rate by 0.48% and 0.82% respectively, and an appropriate increase in the filter bank group can reduce the error rate. In brief the sentence error rate and the word error rate were reduced to 21.06% and 3.12% respectively.