PROC. OF THE SPEECH PROCESSING WORKSHOP, MAGDEBURG, 2003 Various Information Sources for HMM with Weighted Multiple Codebooks

Abstract. When semicontinuous HMM are used for acoustic modeling in speech recognition usually all states share a single codebook. In our investigations we split the feature set into independent subsets and use separate codebooks for each part. This provides a higher modeling flex-ibility while keeping the parameter space compact. Further experiments integrate new information sources by using additional codebooks which are estimated in a supervised training. For instance codebooks for phone transitions are applied. Codebook exponents weight the different infor-mation sources. Relative reductions in word error rate up to 20 % have been achieved.