Laura A. Drake
Janet C. Rutledge
Elec. Eng./Comput. Sci. Dept., Northwestern Univ., Evanston, IL 60208-3118
Spectral representations of speech have been used extensively in speech communication applications such as hearing loss compensation algorithms and speech recognizers. Here, a representation incorporation additional perceptual characteristics is created. It is hoped that such a representation would improve the performance of application systems with noisy input by emphasizing more perceptual cues. For example, the shape of the speech signal is modified in hearing loss compensation to fit the residual spectrum of the listener. Information from perceptual cues relating to the interaction of key components can be used to adjust the compensation algorithm to enhance intelligibility. These same cues that are important to human perception may be incorporated into the feature vector to enhance the performance of a speech recognition system such as an HMM or neural network. Results will be presented at the conference.