Large-vocabulary continuous speech recognition on spontaneous speech task

1. 基于自发语音任务的大词汇连续语音识别在大词汇连续语音识别中，我们研究了几种声学和语言模型的无监督自适应方法，并在自发日语语料库（CSJ）上对这些方法进行了评价。LVCSR系统采用全协方差矩阵作为声学模型。识别实验结果表明，基于词性对自适应数据进行加权后，无监督自适应的词错误率（WER）从无自适应的19.17%下降到14.73%，无监督自适应的词错误率下降到14.47%。并对连续混合FRAM （CHMM）系统和离散混合HMM （DMHMM）系统在CSJ上的性能进行了比较。结果表明，DMHMM系统具有与CHMM系统基本相同的性能，在6000状态24混合DMHMM下，其识别误差率达到19.73%，但目前普遍认为DMHMM的识别错误率要比CHMM高得多。基于离散混合hmm的鲁棒语音识别提出了一种基于离散混合hmm的噪声条件下鲁棒语音识别新方法。dmhmm最初是为了减少解码过程中的计算成本而提出的。近年来，我们将dmhmm应用于噪声语音识别，发现它对噪声语音建模是有效的。为了进一步提高噪声鲁棒性语音识别，提出了一种基于直方图均衡化（HEQ）的dmhmm归一化方法。HEQ方法可以补偿加性噪声的非线性影响。它通常用于连续混合HMM （CHMM）系统的特征空间归一化。本文提出了基于HEQ的dmhmm模型空间归一化和特征空间归一化。在模型空间归一化中，利用HEQ方法导出的变换函数对dmhmm的码本进行修正。将该方法与传统的chmm和dmmm进行了比较。结果表明，采用多变换函数对dmhmm进行模型空间归一化是一种有效的抗噪语音识别方法。少

1. Large-vocabulary continuous speech recognition on spontaneous speech taskIn large-vocabulary continuous speech recognition, we investigate several methods of unsupervised adaptation of both acoustic and language models and evaluate the methods on the Corpus of Spontaneous Japanese (CSJ). The LVCSR system has full-covariance matrices as the acoustic model. The results of recognition experiments showed the decrease in word error rate (WER) from 19.17% without adaptation to 14.73% with unsupervised adaptation, moreover to 14.47% with unsupervised adaptation by weighting the adaptation data on the basis of a part of speech. Also, we compared the performance between continuous-mixture FRAM (CHMM) system and discrete-mixture HMM (DMHMM) system on the CSJ. As a result, DMHMM system provided almost the same performance as the CHMM system and WER of 19.73% had been obtained with 6000-state 24-mixture DMHMMs, though it has been generally believed that the recognition error rates of DMHMM were … More much higher than those of CHMM until now.2. Robust speech recognition using discrete-mixture HMMsWe introduce a new method of robust speech recognition under noisy conditions based on discrete-mixture HMMs (DMHMMs). DMHMMs were originally proposed to reduce calculation costs in the decoding process. Recently, we have applied DMHMMs to noisy speech recognition, and found that they were effective for modeling noisy speech. Towards the further improvement of noise-robust speech recognition, we propose a novel normalization method for DMHMMs based on histogram equalization (HEQ). The HEQ method can compensate the nonlinear effects of additive noise. It is generally used for the feature space normalization of continuous-mixture HMM (CHMM) systems. In this paper, we propose both model space and feature space normalization of DMHMMs by using HEQ. In the model space normalization, codebooks of DMHMMs are modified by the transform function derived from the HEQ method. The proposed method was compared using both conventional CHMMs and DMHMMs. The results showed that the model space normalization of DMHMMs by multiple transform functions was effective for noise-robust speech recognition. Less