Small vocabulary speech recognition system using the Maltese language

Abstract

Over the past few years, speech technologies have evolved dramatically. This has led researchers to start looking for speech technologies which could be used with the Maltese language. As a matter of fact, it is in this spirit which this study is being carried out. The main aim of this study is to develop an automatic speech recognition system (ASR) using the Maltese language. The kind of speech Recognition referred to in this thesis implies a conversion from spoken words or phrases into written text. The speech recognition system implemented in this study has a number of distinctive features. The system is primarily speaker dependent and small vocabulary. This means that the training perfonned is based only on one speaker, therefore the performance is also dependent on that particular speaker while the system is also able to recognize a restricted set of words. Although there are several models that could be used for speech recognition, the Hidden Markov Model (HMM) is used in this implementation. The Hidden Markov Model Toolkit (HTK) is then used to train the monophone HMMs, which are used to perfonn recognition through the Viterbi algorithm. Currently the system is able to provide very promising results, both when used in a speaker dependent environment and also as speaker independent. When tested as a speaker dependent system, using the same speaker used for training, the performance levels are close to 90%. On the other hand when tested using two independent speakers, the results obtained were close to 80% and 67% for male and female speakers respectively. The word error rate is relatively small for a novel speaker dependent system since it is only 15%, however with several improvements these results could also be improved.