Hidden Markov modeling in on-line dysarthric speech recognition

In this paper we report on the development of automatic speech recognition (ASR) system for assistive devices. The target applications are in communication and control devices for persons with severe motor disorders accompanying with severe dysarthria. Using portable PC platform for our experiments, we have developed a adaptive system which performs automatic isolated word detection and word recognition. Mel-frequency cepstral coefficients were used for parametric representation of the speech signal and discrete hidden Markov modeling (DHMM) was used for pattern recognition. The system was tested in one person with cerebral palsy and severe dysarthria. Our initial results demonstrate high generalization over subsequent recording sessions (up to 90%), and lower, but still usable generalization over sessions recorded one week later (76.5%). They also demonstrate improved performance when amount of training data was increased from five to fifteen repetitions.