Development and implementation of speech enhancement techniques
In many speech communication settings, the presence of background interference degrades the quality or intelligibility of speech. In the typical singlechannel speech enhancement systems discussed in this thesis, the noisy speech is picked up by only one microphone. Three techniques - spectral subtraction, self-excited method, and subband structures associated with their corresponding voice activity detector — are developed by using short-term spectral magnitude techniques. The variable-length windows are used in noisy speech component estimates, noise component estimates and voice activity detector. Spectral subtraction and pseudo-quadrature mirror filter (PsQMF) speech enhancement systems are developed based on the TMS320C50 digital signal processor. These systems are designed to remove additive background noise and improve speech quality. After speech enhancement processing, the output is the enhanced speech. Both computer simulations and real-time experiment results are presented to show the performance of these techniques. The Itakura distance was chosen to evaluate the performance of these speech enhancement techniques.