Speech Communication
ICCHP keynote: recognizing silent and weak speech based on electromyography
ICCHP'10 Proceedings of the 12th international conference on Computers helping people with special needs: Part I
| Hi-index | 0.00 |
Automatic speech recognition (ASR) is an alternative control methodology being investigated for high performance jet aircraft; however, performance of conventional acoustic ASR systems is severely degraded by various noise and stress conditions during flight. Myoelectric signals (MESs) from articulatory muscles of the face are proposed as a secondary source of speech information to enhance conventional ASR systems. Employing a MES classification method previously used for prosthetic control, MES ASR classification accuracies above 90% are obtained for a ten word vocabulary. This surpasses previous attempts at MES ASR, which had classification accuracies below 65% for ten word vocabularies. A new method of MES ASR, using hidden Markov models (HMMs) is introduced. This classification technique is demonstrated to have a greater resilience to temporal variance, which is important for ASR. A novel method of combining the opinion of experts, called the plausibility method, is introduced. The plausibility method is founded in evidence theory, which is a mathematical framework that enables the precise assignment of partial beliefs and the ability of combining partial beliefs from multiple bodies of evidence. Using the plausibility method, the MES ASR expert is combined with an acoustic ASR expert to form a multi-expert ASR system. This multi-expert ASR system is evaluated across multiple levels of acoustic noise and positive pressure, and compared to other methods of combining the opinion experts, including the Borda count method and a score-based method. The plausibility method demonstrates higher classification accuracies than other methods of combining the opinion of experts. Also, the plausibility method is also able to dynamically track the reliability of an expert without monitoring the external environment. It is also demonstrated that the addition of another MES ASR expert further increases classification accuracy. The advantages of classifying MES using HMMs and the benefits of a multi-expert system are also shown in prosthetic control. HMM classification of MES has a higher consistency in classification because decisions are based upon previous decisions and this consistency improves the overall classification accuracy. Further improvements in classification accuracy are obtained using a multi-expert system.