The application of continuous state HMMs to an automatic speech recognition task

Seivwright, Chloe (2023). The application of continuous state HMMs to an automatic speech recognition task. University of Birmingham. Ph.D.

Preview

Seivwright2023PhD.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (18MB) | Preview

Abstract

Hidden Markov Models (HMMs) have been a popular choice for automatic speech recognition (ASR) for several decades due to their mathematical formulation and computational efficiency, which has consistently resulted in a better performance compared to other methods during this period. However, HMMs are based on the assumption of statistical independence among speech frames, which conflicts with the physiological basis of speech production. Consequently, researchers have produced a substantial amount of literature to extend the HMM model assumptions and incorporate dynamic properties of speech into the underlying model. One such approach involves segmental models, which addresses a frame-wise independence assumption. However, the computational inefficiencies associated with segmental models have limited their practical application. In recent years, there has been a shift from HMM-based systems to neural networks (NN) and deep learning approaches, which offer superior performance com- pared to conventional statistical models. However, as the complexity of neural models increases, so does the number of parameters involved, requiring a greater dependency on training data to optimise model parameters.

This present study extends prior research on segmental HMMs by introducing a Segmental Continuous-State Hidden Markov Model (CSHMM) examining a resolution to the issue of inter-segmental continuity. This is an alternative approach when compared to contemporary speech modelling methods that rely on data-centric NN techniques, with the goal of establishing a statistical model that more accurately reflects the speech production process. The Continuous-State Segmental model offers a flexible mathematical framework which can impose a continuity constraint between adjoining segments addressing a fundamental drawback of conventional HMMs, namely, the independence assumption. Additionally, the CSHMM also benefits from a practical training and decoding algorithm which overcomes the computational inefficiency inherent in conventional decoding algorithms for traditional Segmental HMMs.

This study has formulated four trajectory-based segmental models using a CSHMM framework. CSHMMs have not been extensively studied for ASR tasks due to the absence of open-source standardised speech tool-kits that enable convenient exploration of CSHMMs. As a result, to perform sufficient experiments in this study, training and decoding software has been developed, which can be accessed in (Seivwright, 2015).

The experiments in this study report baseline phone recognition results for the four distinct Segmental CSHMM systems using the TIMIT database. These baseline results are compared against a simple Hidden Markov Model-Gaussian Mixture Model (HMM- GMM) system. In all experiments, a compact acoustic feature representation in the form of bottleneck features (BNF), is employed, motivated by an investigation into the BNFs and their relationship to articulatory properties. Although the proposed CSHMM systems do not surpass discrete-state HMMs in performance, this research has demonstrated a strong association between inter-segmental continuity and the corresponding phonetic categories being modelled. Furthermore, this thesis presents a method for achieving finer control over continuity between segments, which can be expanded to investigate co-articulation in the context of CSHMMs.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):