Automatic detection of speech disorders with the use of Hidden Markov Model

The most frequently used methods of automatic detection and classification of speech disorders are based on experimental determination of specific distinctive features for a given kind of disorder, and working out a suitable algorithm that finds such a disorder in the acoustic signal. For example, for detection of prolonged phonemes, analysis of the duration of articulation is used, and on the contrary, phoneme repetition can be detected with the spectrum correlation methods. Additionally, in the case of prolonged phonemes, classification based on their kind is required (nasal or whispered phonemes, vowels, consonants, etc). Therefore, for every kind of a disorder, a separate algorithm needs to be worked out. Another, more flexible approach is the application of the Hidden Markov Models (HMM). For the needs of the presented work, the HMM procedures were implemented and some basic tests of speech disorder detection were conducted.