Speech Recognition by Human and Machine

Abstract

Several feature extraction techniques, algorithms and toolkits are researched to investigate how speech recognition is performed. Spectrograms were found to be the simplest feature extraction techniques for visual representation of speech, and are explored and experimented with to see how phonemes are recognized. Hidden Markov models were found to be the best algorithms used for speech recognition. Hidden Markov model toolkit and Center for Spoken Language Understanding Toolkit, which are based on hidden Markov models, were not found to be suitable for the intentions of the thesis.