Robust HI and dysarthric speaker recognition - perceptual features and models

This paper explores the necessity of having hearing impaired (HI) and dysarthric speakers be part of the person authentication system and it is considered to be imperative. Automated system on identifying speakers is evaluated by having the perceptual features with critical band analysis done in various non-linear frequency scales and vector quantization (VQ) & Fuzzy C means (FCM) based iterative clustering templates and multi-variant hidden Markov (MHMM) models as representative of HI or dysarthric speakers. For developing a training system, perceptual features are extracted from the speeches of HI or dysarthric speakers after the initial pre-processing techniques namely voice activity detection, pre-emphasis, frame blocking, and windowing contemplated on the speech utterances, and VQ & FCM clustering models and MHMM models are created for each speaker and the study is done on varying cluster and mixture size. The testing phase emphasizes the extraction of features from the test utterances, application of features to the templates, and classification is done based on minimum distance criterion for clustering technique and maximum log-likelihood criterion for MHMM technique. This algorithm gives the overall accuracy of 100% when the decision level fusion classification is done for the perceptual features with critical band analysis done in MEL, BARK, and ERB scales for all the clusters with variations in cluster size for both hearing impaired and dysarthric speaker recognition. Decision level fusion classification using FCM and MHMM technique provides low overall accuracy as compared to the VQ technique.