Improving hidden Markov models for speech recognition

Carl Dennis Mitchell, Purdue University

Abstract

Despite the success of hidden Markov models, there are several assumptions implied by the standard model that are inappropriate for speech. For example, the implied geometric state duration is seldom appropriate for speech modeling. Much of the research in this dissertation has focused on relaxing the assumptions that are inappropriate for speech modeling. Additionally, implementation issues have been addressed. A new method of duration modeling is introduced for a hidden Markov model (HMM). The nonparametric explicit duration model proposed by Ferguson has been extended to any probability mass function or probability density in the exponential family. This model allows for smoother and more accurate duration modeling. Duration modeling is computationally expensive. In order to facilitate practical experimentation, the explicit duration model has been developed on a MasPar MP1, a massively parallel SIMD machine. Additionally, new recursions are introduced which reduce complexity by approximately an order of magnitude for the case of continuous output HMMs. The lower complexity applies to both serial and parallel implementations of explicit duration HMMs. A new method of integrating external segmentation information into the HMM paradigm is proposed. It is shown that HMM phone recognition accuracy can be increased by utilizing an estimate of spectral variation. In addition to the improvements for acoustic modeling, the problem of going from phone recognition to word recognition is addressed. A new HMM variation called the stochastic output HMM (SOHMM) is proposed. The HMM paradigm has been modified to process probability distributions instead of observations. The SOHMM has demonstrated superior performance for high perplexity continuous word recognition.

Degree

Ph.D.

Advisors

Harper, Purdue University.

Subject Area

Electrical engineering|Linguistics

COinS