HMM-based speech recognition system for the dysarthric speech evaluation of articulatory subsystem

Dysarthria is a neuromotor impairment of speech that affects one or more of the speech subsystems, but is often associated with irregular co-ordination of articulators and restricted movement of articulators among other problems. It is reflected in the acoustic characteristics of the phonemes as deviations from their healthy counterparts. To capture these deviations, in this work, isolated-style, phoneme recognition system is developed using monophone as the sub word unit. The performance of this phoneme recognition system for a dysarthric speaker can be directly related to the severity of the problem. To train the sub word unit models, speech data is collected from seven normal speakers. Time-aligned phonetic transcriptions are derived using forced Viterbi alignment procedure. Using this data, hidden Markov models for the required phonemes are trained. Nemours database contains time-aligned phonetic transcriptions for all the ten dysarthric speakers. Using these transcriptions, phonetic inventory is created for each of the dysarthric speakers separately. These phoneme segments are tested with the phoneme models trained using the normal speakers´ data. The performance of this speech recognition system is analyzed after phoneme grouping, based on the place of articulation, for the assessment of the articulatory subsystem of the dysarthric speech. The analysis output correlates well with the Frenchey dysarthria assessment (FDA) scores provided with the database.