Hitoshi Honda

Stochastic features for noise robust speech recognition

This paper describes a novel technique for noise robust speech recognition, which can incorporate the characteristics of noise distribution directly in features. The feature itself of each analysis frame has a stochastic form, which can represent the probability density function of the estimated speech component in the noisy speech. Using the sequence of the probability density functions of the estimated speech components and hidden Markov modelling of clean speech, the observation probability of the noisy speech is calculated. In the whole process of the technique, the explicit information on the SNR is not used. The technique is evaluated by large vocabulary isolated word recognition under car noise environment, and is found to have clearly outperformed nonlinear spectral subtraction (with between 13% and 44% reduction in recognition errors).

Development of the compact English LVCSR acoustic model for embedded entertainment robot applications

In this paper we discuss two techniques to reduce the size of the acoustic model while maintaining or improving the accuracy of the recognition engine. The first technique, demiphone modeling, tries to reduce the redundancy existing in a context dependent state-clustered Hidden Markov Model (HMM). Three-state demiphones optimally designed from the triphone decision tree are introduced to drastically reduce the phone space of the acoustic model and to improve system accuracy. The second redundancy elimination technique is a more classical approach based on parameter tying. Similar vectors of variances in each HMM cluster are tied together to reduce the number of parameters. The closeness between the vectors of variances is measured using a Vector Quantizer (VQ) to maintain the information provided by the variances parameters. The paper also reports speech recognition improvements using assignment of variable number Gaussians per cluster and gender-based HMMs. The main motivation behind these techniques is to improve the acoustic model and at the same time lower its memory usage. These techniques may help in reducing memory and improving accuracy of an embedded Large Vocabulary Continuous Speech Recognition (LVCSR) application.