Hybrid model based approaches for dysarthric speech recognition

Abstract

Automatic Speech Recognition ASR involves conversion of speech signal into text The automatic speech recognition systems are used in various fields such as voice dialing call routing health care military and robotics In spite of the advances in speech technology their benefits have not been available to persons suffering from dysarthria a kind of motor speech disorder caused by neurological injury to the central or the peripheral nervous system In dysarthric persons speech production subsystems such as respiration phonation resonance prosody and articulation can be affected leading to the detriments in intelligibility audibility naturalness and potency of vocal communication This kind of disorder is caused by a stroke muscular dystrophy brain injury tumor Parkinsons disease Huntingtons disease or multiple sclerosis Some dysarthric speech characteristics are mono pitch harsh voice vowel distortions and strained strangled vocal quality Due to these characteristics the pronunciation often suffers from the following limitations the rate of the dysarthric speech is lower there is no consistency in pronunciation pronunciation varies due to fatigue speaking rate is slow Developing ASR systems specifically designed for people suffering from dysarthria can be very helpful In this thesis we explore approaches to recognize dysarthric speech patterns Hidden Markov Model HMM is a technique for probabilistic modeling of sequential data HMMs have proved to be effective in various applications such as speech recognition time series modeling and gesture recognition In this method the sequence of feature vectors is extracted from the speech utterance of a sub word unit or a word unit newline