Development of Noise Robust Speech Recognition and Its Application on Mobile Environment

本文提出了基于离散混合hmm （dmhmm）的鲁棒语音识别新方法。这项工作的目的是为包含平稳和非平稳噪声的不利条件开发鲁棒语音识别。我们特别关注了脉冲噪声问题，这是实际语音识别系统中的一个主要问题。为了解决这个问题，我们提出了两种方法。首先，提出了一种基于MAP的DMHMM参数估计方法，以提高可训练性。第二种方法是用阈值补偿dmhmm的观测概率，以减少异常值的不利影响。对日语读报语音的LVCSR实验评估表明，在同时包含平稳噪声和脉冲噪声的不利条件下，该方法的平均错误率降低了28.1%。2)基于模型直方图均衡化的dmhmm噪声鲁棒语音识别为了进一步改进噪声语音识别，本文提出了一种新的dmhmm码本归一化方法。码本归一化方法基于直方图均衡化（HEQ），可以补偿模型空间中加性噪声的非线性影响。将该方法与传统连续混合hmm （chmm）和dmhmm进行了比较。结果表明，所提出的方法获得了最佳的性能，比CHMM基线平均相对提高了29.2%。

1) Noisy speech recognition using DMHMMsWe have proposed new methods of robust speech recognition using discrete-mixture HMMs (DMHMMs). The aim of this work is to develop robust speech recognition for adverse conditions that contain both stationary and non-stationary noise. In particular, we focus on the issue of impulsive noise, which is a major problem in practical speech recognition system. In order to solve the problem, we have proposed two methods. First, an estimation method of DMHMM parameters based on MAP has been proposed aiming to improve trainability. The second is a method of compensating the observation probabilities of DMHMMs by threshold to reduce adverse effect of outlier values. Experimental evaluations on Japanese LVCSR for read newspaper speech showed that the proposed method achieved the average error rate reduction of 28.1% in adverse conditions that contain both stationary and impulsive noises.2) Model Based Histogram Equalization for Noise Robust Speech Recognition by Using DMHMMsTowards further improvement of noisy speech recognition, we have proposed a novel normalization method for codebooks of DMHMMs in this paper. The codebook normalization method is based on histogram equalization (HEQ) and it can compensate the non-linear effects of additive noise in model space. The proposed method was compared with both conventional continuous-mixture HMMs (CHMMs) and DMHMMs. It showed that the proposed method obtained the best performance, and obtained an average relative improvement of 29.2% over the CHMM baseline.