Improved Hidden Markov Modeling for Speaker-Independent Continuous Speech Recognition

Improved Hidden Markov Modeling for Speaker-Independent Continuous Speech Recognition Xuedong Huang, Fil Alleva, Satoru Hayamizu Hsiao-Wuen Hon, Mei-Yuh Hwang, Kai-Fu Lee School of Computer Science Carnegie Mellon University Pittsburgh, Pennsylvania 15213 Abstract This paper reports recent efforts to further improve the perfor- mance of the Sphinx system for speaker-independent contin- uous speech recognition. The recognition error rate is signifi- cantly reduced with incorporation of additional dynamic fea- tures, semi-continuous hidden Markov models, and speaker clustering. For the June 1990 (RM2) evaluation test set, the error rates of our current system are 4.3% and 19.9% for word-pair grammar and no grammar respectively. Introduction This paper reports our recent effort to further improve the accuracy of the Sphinx System [10]. We choose here to adhere to the basic architecture of the Sphinx System, and to use the standard Resource Management task and training corpus. Possible improvements could be resulted from the following categories: • Incorporate additional dynamic features. • Improve HMM probabilistic representation. • Cluster training speakers to provide multiple models. • Introduce discfiminant transformations to improve dis- crimination. • Extend corrective training for semi-continuous models. • Improve allophonetic models of coarticulation. This paper will report advances on the first five categories. Improved allophonic modeling has been reported in [13]. Our first improvement involves the incorporation of dy- namic features computed from the LPC cepstrum. Previous versions of the Sphinx system have used first order differ- enced cepstrum and power. Here, we experimented with second and third order differenced cepstrum and power. We also experimented with incorporation of both 40 msec and 80 msec differenced cepstrum, as well as the difference derived from compressed speech [3]. These additional feature sets are incorporated in the multi-codebook framework. The best combination reduced errors by 17% over the baseline results. Our second technique employs the semi-continuous hid- den Markov model (HMM) [8]. Multiple-codebook semi- continuous models are extended to our current Sphinx ver- sion. Both diagonal and full covariance Gaussian models are investigated. We found that the best variants of both models reduced error rate of discrete HMMs by 10-20%. Due to smoothing abilities of the semi-continuous model, we were able to train multiple sets of models for different speakers. We investigated automatic speaker clustering and explicit male/female clustered models. In both cases, models of all the speaker clusters are simultaneously active, with the restriction that no between-cluster transitions are allowed. Thus, the system retains speaker-independent characteristics. By using multiple model sets with the semi-continuous HMM, the error rate is further reduced by 10-15%. We experimented with two variants of linear discriminant transformations. The first attempted to use a single transfor- marion to separate all triphone states. The second attempted to shift the mean vectors of the semi-continuous mixtures, so as to separate confusable words. However, neither method produced any improvement. Finally, we investigated corrective training for semi- continuous models. At the time of this writing, we have only applied our discrete corrective training algorithm [15] to semi-continuous models. We found that this method is …

Content truncated to 500 words