Maltese speech recognition over mobile telephony

Abstract

Automatic speech recognition is the technology that allows humans to communicate to a machine through spoken utterances instead of the more conventional means such as keyboards. The spoken signal is captured and analysed in a series of one or more steps, and action is taken depending on the final representation. Speech recognition inherits much from the cognitive sciences since most approaches try to first understand how humans perceive and interpret speech, and then build a model that best represents human perception. While models for high performance speech recognition have been studied extensively for many decades, it is only recently that we have seen some impressive results. The main reason for this is that unfeasible models for high performance systems designed in the past have become feasible due to the increase in the available processing power. Modern speech recognisers extract features from the time-varying signal that discriminate between the different units of speech, being affected only slighty by the pitch or manner of the speaker. The features are then passed through one or more trained Hidden Markov Models (HMMs) for the classification of a phonetic sequence, and ultimately the utterance itself [18, 13].