A tutorial on using hidden Markov models for phoneme recognition

A phoneme is a contrastive unit in the sound system of a particular language that helps us distinguish between meanings of words from a set of similar sounds corresponding to it pronounced in one or more ways, depending on the configuration of the articulators like air flow, teeth, tongue, lip and vocal chord movements. In this paper we explain in detail the process of phoneme recognition using hidden Markov models (HMM), implemented with HTK (hidden Markov model toolkit) developed by Cambridge University. An HMM is a statistical model that describes a probability distribution over a number of possible sequences. HTK is widely used in speech recognition research worldwide and aims at manipulating and building hidden Markov models. This paper follows a 3-state HMM to give an insight into the training and testing phases of phoneme recognition. We were able to achieve up to 100% accuracy in recognizing the phonemes, and thus strongly believe that a 3-state HMM ideally suits our requirements. We also explore the features of various accents, evaluating their cepstral coefficients and perform time and frequency domain analysis.