Automated assessment and treatment of speech rate and intonation in dysarthria
Proceedings of the 7th International Conference on Pervasive Computing Technologies for Healthcare
| Hi-index | 0.00 |
Acoustic analysis of speech is gaining popularity for diagnosing speech disorders. Research is always carried out to develop objective measures that correlate well with perceptual judgments to assess speech characteristics. Speech rate is one such characteristic that is often studied to assess a medical condition in a subject. Speech rate also has applications in automatic speech recognition (ASR) systems as a very high or very low speech rate reduces the system accuracy. A novel hidden Markov model (HMM) based method to count the number of syllables in continuous speech has been developed in this research. This number of syllables can then be used to estimate speech rate in syllables/second. The method has been tested on normal and dysarthric speech. For normal, an error of about 15% is obtained, which is better than presently available techniques, while for dysarthric subjects, the error was 27%. The approach was also tested on diadochokinetic speech of dysarthric subjects, which includes repetitions of one-to-three syllables. An error of 6.9% is obtained with discrete HMMs. Another aspect of this research is the design of a novel dysarthria classifier using global statistics of speech. Dysarthria is a class of neurological disorders that hampers speech production. Depending on the location of lesion in the nervous system, dysarthria can be classified into five main types. Clinicians find it useful to identify the type of dysarthria of a subject to determine the location of injury and to structure the treatment and therapy required. There is evidence suggesting a cause-effect relationship between the dysarthria type and the resulting speech characteristics. Therefore, the observed speech characteristics in a subject can be analyzed to identify the type of dysarthria. The dysarthria classifier designed in this research uses global statistics (mean, variance, etc.) of features such as Mel-frequency cepstral coefficients, perceptual linear prediction, etc., as inputs. An error of 10.5% is achieved for a tree-like classifier that distinguishes between amyotrophic lateral sclerosis (ALS), flaccid dysarthria (FD) and spastic dysarthria (SD). The performance of the classifier is compared with some commercially available software for making decision trees.