Speech recognition systems commonly use a beads-on-string approach in decoding the input signal. They attempt to subdivide the input signal and recongise subword units known as phonemes. To aid this recognition process, a word ngram (typically bigram or trigram) language model is used to help predict what the next word will be given the n-1 previous words - the words are converted to a sequence of phonemes using a pronunciation dictionary.
This project aims to improve language modelling, and thus speech recognition, by modelling formulaic language. Formulaic language, produced by humans, is retrieved whole from memory, and undergoes no novel construction. For example, typically in conversational speech, a human will use common phrases and idioms, such as such as "kick the bucket", "at the end of the day", in an attempt achieve maximum fluency. These phrases are retrieved quickly from memory, and are thus a more efficient method for speech production in a conversational situation.
Mr. Christopher Watkins, Prof. Stephen Cox