| The thesis introduces an enhanced phoneme-based myoelectric speech recognition system. The system can recognize new words without retraining the phoneme classifier, which is considered as the main advantage of phoneme-based speech recognition. However, as more words were added to the test set, the performance of the prior system (E. J. Scheme 2005) decreased when recognizing speech using information from the facial myoelectric signal (MES). To maintain a reasonable recognition result, several improvements are proposed. For the current MES speech recognition approach, the raw MES is processed by class-specific rotation matrices to spatially decorrelate the data prior to feature extraction in a pre-processing stage. Then an uncorrelated linear discriminant analysis (ULDA) is used for dimensionality reduction. The resulting data are classified through an HMM classifier to obtain the phonemic log likelihoods, which are mapped to corresponding words using a word classifier. An average word classification accuracy of 98.533% is achieved over six subjects. |
