Improving Speech Classification Accuracy: A Support Vector Machine Approach (#1554)

Authors

Thorpe, Balvin

Abstract

Digital hearing aids improve hearing by reducing background noise and significantly improve sound qualities. However, digital hearing aids do not produce significant improvements for the hearing impaired, when the challenge for some hearing-impaired persons, is recognizing high frequency sounds, such as consonants. Digital hearing aids (DHAs) use parametric phoneme classifiers, such as Hidden Markov Model (HMM). These classifiers produce at best 80% phoneme classification accuracy. The main research question therefore was: - Could the digital hearing aid’s phoneme classification accuracy improve, if a non-statistical/non-parametric phoneme classification algorithm was developed and used in the processor of the digital hearing aid? In this study, speech was classified using linear Support Vector Machine (SVM), a non-parametric classifier, to identify vowels differently from consonants. SVM was chosen, as literature indicate that SVM non-parametric classifier was likely to return the highest classification accuracy among all classifier algorithms. The SVM classifier was built in MATLAB by parsing the phonemes from Texas Instrument and Massachusetts Institute of Technology (TIMIT) training speech files, and generating the corresponding Mel Frequency Cepstrum Coefficients (MFCC) for each phoneme. The built SVM Classifier was tested, using files from the TIMIT speech test database. Results showed that the built SVM classifier produced phoneme classification accuracy ranging from 74% to 92.7%. These results indicate that the built SVM can be used to classify phonemes with, accuracy that is equal to or better that the existing statistical/parametric phoneme classifiers.