Finite state models for recognition and validation of read prompts

An automatic speech recognition system has to combine acoustic and linguistic information. Therefore the search space spans multiple layers. Finite state models and weighted finite state transducers in particular can efficiently represent this search space by modeling each layer as a transducer and combining them using generic weighted finite state transducer algorithms. When recognising a text prompt being read aloud, the prompt gives a good estimate of what is going to be said. However human reading naturally produces some deviations from the text, called miscues. The purpose of this thesis is to create a system which accurately recognises recordings of reading. A miscue tolerant finite state language model is implemented and compared against two traditional approaches, an N-gram model and forced alignment.

The recognition result will ultimately be used to validate the recording as fit for further automatic processing in a spoken foreign language exam, which Project DigiTala is designing for the Finnish matriculation examination. The computerization of the matriculation examination in Finland makes the use of such automatic tools possible.

This thesis first introduces the context for the task of recognising and validating reading. Then it explores three methodologies needed to solve the task: automatic speech recognition, finite state models, and the modeling of reading. Next it recounts the implementation of the miscue tolerant finite state language models and the two baseline methods. After that it describes experiments which show that the miscue tolerant finite state language models solve the task of this thesis significantly better than the baseline methods. Finally the thesis concludes with a discussion of the results and future work.