Adaptive model-based speech enhancement

Declaration This dissertation is the result of my own work and includes nothing which is the outcome of work done in collaboration except where stated. It has not been submitted in whole or part for a degree at any other university. The length of this thesis including footnotes and appendices is approximately 37000 words. i Summary This dissertation details the development and evaluation of techniques to enhance speech corrupted by unknown independent additive noise when only a single microphone is available. It therefore seeks to address a deeciency of many speech enhancement systems which require a priori knowledge of the interfering noise statistics. Such a deeciency must be corrected if these systems are to operate in real world situations. The enhancement systems developed are based on an existing system by Ephraim (Ephraim 1992a). This approach models the speech and noise statistics using autoregressive hidden Markov models (AR-HMMs). Two main extensions to this technique are developed in order to make it adap-tive. The rst estimates the noise statistics from detected pauses. The second forms maximum likelihood estimates of the unknown noise parameters using the whole utterance. Both techniques operate within the AR-HMM framework. Additional work in this dissertation improves the modelling power of AR-HMM systems by incorporating perceptual frequency. The bilinear transform is used to warp the frequency spectrum of the feature vectors to an approximation of the Bark scale. This modiication can be incorporated into both AR-HMM recognition and enhancement systems. The enhancement techniques are evaluated on the NOISEX-92 and Resource Management (RM) databases, giving indications of performance on simple and more complex tasks respectively. Additional experiments investigating the incorporation of perceptual frequency into AR-HMM systems were conducted on the E-set of the speaker independent ISOLET database. Both enhancement schemes proposed were able to improve substantially on baseline results. The technique of forming maximum likelihood estimates of the noise parameters was found to be the most eeective. Its performance was evaluated over a wide range of noise conditions ranging from-6dB to 18dB and on various types of stationary real-world noises. The incorporation of perceptual frequency into AR-HMM systems was found to increase recognition performance substantially on both the ISO-LET and RM databases. The improvement was less marked for the more complex task, highlighting that AR-HMMs could beneet from the inclusion of more variance information. Acknowledgements First I would like to thank my supervisor Tony Robinson. He provided me with the …