Combining Neural Networks And Hidden Markov Models For Continuous Speech Recognition

Pure MLP-based approaches have not previously been t demonstrated to function well for continuous-speech recogni ion because of the need for accurate segmentation of the w speech signal. HMMs, on the other hand, provide a frame ork for simultaneous segmentation and classification of t speech, which has been demonstrated to be useful for con inuous recognition. Previous work by Morgan and Bourlard d H [1] has shown theoretically and practically that MLPs an MMs can be combined by using MLPs for the estimation , t of the HMM state-dependent observation probabilities hereby exploiting the advantages of both approaches.