Nonparametric hidden Markov models: Principles and applications to speech recognition

Continuous-density hidden Markov models (HMM) are a popular approach to the problem of modeling sequential data, e.g. in automatic speech recognition (ASR), off-line handwritten text recognition, and bioinformatics. HMMs rely on strong assumptions on their statistical properties, e.g. the arbitrary parametric assumption on the form of the emission probability density functions (pdfs). This chapter proposes a nonparametric HMM based on connectionist estimates of the emission pdfs, featuring a global gradient-ascent training algorithm over the maximum-likelihood criterion. Robustness to noise may be further increased relying on a soft parameter grouping technique, namely the introduction of adaptive amplitudes of activation functions. Applications to ASR tasks are presented and analyzed, evaluating the behavior of the proposed paradigm and allowing for a comparison with standard HMMs with Gaussian mixtures, as well as with other state-of-the-art neural net/HMM hybrids.