Robust estimation of HMM parameters using fuzzy vector quantization and Parzen's window

Abstract The paper presents a new parameter estimation method for discrete hidden Markov models (HMM) in speech recognition. This method makes use of fuzzy vector quantization and the Parzen window in dealing with the problem of insufficient training data, and it therefore may be regarded as a kind of maximum likelihood estimation with some smoothing effect. The proposed method is compared with the existing smoothing techniques by experiments of speaker-independent isolated word recognition. The results show that the new estimation method has led to improved recognition and therefore it may be used as an alternative to the parameter smoothing techniques for HMMs.