Hidden Markov model toolkit for speech recognition using MATLAB

Abstract

The main purpose of the study was to develop a speech recognition tool using Hidden Markov Model techniques for the Maltese language. Speech recognition of Maltese is an area that needs lots of investigation since few researchers studied the topic vis-a-vis the Maltese language. The products of the study include a Maltese phoneme and utterance speech recogniser, a recipe for building Hidden Markov Model speech recognisers, especially for the Maltese language. Phonetically and labelled corpora are very important and useful in building a speech recognition system. Two different corpora were collected of audio recordings in Maltese, in which subjects read aloud whole sentences. The first type of corpora contained the training data and the other the testing data. The project was implemented using MATLAB. Signal processing in the time and frequency domain helps in the implementation of the system. MATLAB straight forward and easy programming interface in conjunction with its built in functions for frequency domain analysis makes it an ideal tool for speech analysis projects. Several algorithms and toolboxes exist for building speech recognition tools. One of the most efficient methods is the Hidden Markov Model, since it gives the best results when compared to others. This is done in two stages: the training and the recognition or decoding phase. In the first stages of the system, i.e. the training stage data is prepared in a format that will be suitable to work with, such as dividing data into frames and so on. Afterwards, Hidden Markov Models are created and initialised for each phoneme. The toolbox within the MATLAB environment provides the use of conventional techniques for the continuous and discrete environments. Also the different groups of vector patterns are provided. After the model is trained, several re -estimation algorithms such as the Baum-Welch and Viterbi are used so that each HMM represents its corresponding utterance. In the second stage of the system, the recognition stage, testing utterances are used to evaluate the performance of the system. In other words, test data is used and the Hidden Markov Model generated will be tested against this data. If the model is trained correctly, the test data will be recognised. The developed system can be used by developers and researchers interested in speech recognition for the Maltese language and thus the study could cater for large vocabularies of this language