Speech recognition using hidden markov model with low redundancy in the observation space

Abstract

Current speech recognition systems usually model a speech signal as a finite-state stochastic process, in which acoustic observations are obtained through short-term spectral analysis. The model has to deal with several thousands of speech parameters during one second of utterance. A great redundancy in the parameters makes processing computationally very expensive. We propose a combination of 2-D cepstral analysis and continuous Hidden Markov Model with a small, optimally designed, number of states and acoustic observations. 2-D cepstrum efficiently preserves spectral variations of speech and yields uncorrelated parameters in both time and frequency. The system is evaluated on isolated word recognition task in Slovak language. Promising preliminary results are presented.