本文提出一种新的描述正常语音和变异语音之间关系的补偿因子。该补偿因子兼顾考虑了由于变异引起的特征分布中均值和方差的变化，并在k-均值初始化的参数基础上，采用EM 算法迭代估计变异补偿因子的值。通过估计出的补偿因子对变异语音特征进行补偿。对航空模拟飞行器中采集的应力变异下特定话者小词表孤立词的实验结果表明，利用所提出的方法可以将识别率提高32.3%。

本文从三个不同的角度分析了稳健水印应满足的性质，提出了相应的三个原则。对于不满足原则的水印给出了相应的转化方法，从而使处理不同类型的水印有了一个统一方法，并进行了相关实验。结果表明，在遵循上述原则的情况下，可以极大地提高水印的抗攻击能力。

文讨论了变异语音处理技术及其研究进展，分析了变异情况对语音识别性能产生的影响，综述了变异语音分类和变异语音识别方法，探讨了变异语音处理研究中存在的问题及未来的研究重点。

本文介绍了基于语音信号的情感处理技术的研究进展，综述了该领域的研究方法、相关技术及应用领域，讨论了其未来的发展方向。

在语言发音学习中，有效的反馈对学习者有很大的帮助。计算机辅助发音学习系统可以给学习者有效的发音指导。本文就目前基于语音识别的发音学习技术进行介绍，给出系统原理框图，对一些关键技术和问题进行探讨，最后对其发展进行展望。

Speech production variations due to perceptually induced stress contribute significantly to reduced speech processing performance. One approach that can improve the robustness of speech recognition against stress is to formulate an objective classification of speaker stress based on the acoustic speech signal. Special processing could then be applied once non-neutral stress states are detected. Thus, it is very important to study an effective classification method of speech under stress. So far, there are little works about the conditions of G-Force in the studies of stressed speech. In this paper, we investigated the speech features of pitch and TEO pitch for classification of neutral speech and speech under G-Force. Both Bayesian hypothesis and HMM classifier are employed for stress classification. Experimental results show that TEO pitch is better than pitch for classification of speech under G-Force and HMM classifier is also better than Bayesian classifier. For HMM classifier, using both TEO pitch and its delta feature is better than just using TEO pitch, 89.2% and 97.3% classification rates are gotten for two speakers, respectively.

It is well known that differences between training and testing environments seriously affect speech recognition accuracy. Several environment adaptation techniques have been proposed for eliminating environmental difference, and they are effective when the adaptation data in the testing environment is available beforehand. However, testing environments are not always known beforehand. Therefore, a new framework, which uses testing utterances themselves for adaptation, is effective to cope with such variation. In this paper, a method of testing environment adaptation is proposed, which radually learns the features of testing environment as the testing procedure, and does not need to get some testing samples beforehand.

Learning the influence of additive noise and channel distortions from training data is an effective approach for robust speech recognition. We had proposed a novel method of discriminative learning environmental features according to Minimum Classification Error (MCE) criterion in the previous work, in which additive noise are expressed by the weighted combination of multiple types of noises, and the channel distortions are assumed to be consisted of the channel distortions of the whole training data and the current utterance. In this paper, we use a Gaussian distribution to stand for the distribution of additive noise, and adaptive learn the combination factors of the channel distortions. The current method has been proved better than the former one by experiments.

Learning the influence of additive noise and channel distortions from training data is an effective approach for robust speech recognition. Most of the previous methods are based on maximum likelihood estimation criterion. In this paper, we propose a new method of discriminative learning environmental parameters, which is based on Minimum Classification Error (MCE) criterion. By using a simple classifier defined by ourselves and the Generalized Probabilistic Descent (GPD) algorithm, we iteratively learn environmental parameters. After getting the parameters, we estimate the clean speech features from the observed speech features and then use the estimation of the clean speech features to train or test the back-end HMM classifier. The best error rate reduction of 32.1% is obtained, tested on a Korean 18 isolated confusion words task, relative to conventional HMM system.

Channel compensation technique has been proved to be an effective approach for robust speech recognition. In this paper, we compare the performance of our proposed method RMFCC with those of the former channel compensation methods: CMS, two-level CMS and RASTA for robust telephone speech recognition. For all experiments, a Korean isolated 84-word-database consisting of 80 speakers collected from local telephone line is adopted. Using RMFCC, a 39.8% reduction in word error rate is obtained relative to conventionalHMMsystem. It is shown from the experiments that RMFCC, comparing with RASTA, reduces the computational complexity without losing accuracy, and is also better than CMS and two-level CMS on the performance. After discussion, we verify that it is an effective approach to suppress very low modulation frequencies by fjiltering for robust telephone speech recognition.