Vinsamlegast notið þetta auðkenni þegar þið vitnið til verksins eða tengið í það: http://hdl.handle.net/1946/31280
In the past few years, automatic speech recognition (ASR) has become an important com-ponent in human-computer interaction. It is crucial for less-resourced languages, such as Icelandic, that effort is put into developing ASR systems for the languages to maintain their use in the technology advances of our time. In this project an open ASR recipe, developed at Reykjavik University, is used to train an ASR for Icelandic. The data used to train the ASR are Althingi’s parliamentary speeches. An ASR has already been implemented for the speeches with good results, but the ASR built in this project, is trained on only a subset of the speeches, and used as a test ASR for the lattice re-scoring presented in the project. The lattice re-scoring method, finds the new best-scoring paths for incorrectly transcribed utterances, based on manual error corrections. Even the best ASRs make mistakes that will completely change the meaning of the recognition result. The results of these systems are therefore often reviewed by a human editor, that fixes the errors that arise. Offering automatic updates of utterances, with lattice re-scoring, could decrease the manual labor needed to fix the errors.
This project investigates whether fixing the first error will put the ASR on the right track so that later errors that occur, will automatically be fixed. The new path is found by manually correcting the first error of the utterance and computing a new path through the correction, using the lattices created during the ASR decoding process. The method shows promising results. The sentence error rate (SER), for utterances with two errors, drops to 82.77%, and for utterances containing three errors it decreases to 95.88%. This indicates that the
re-scoring method is able to automatically fix all the errors, which follow the correction, in a proportion of the utterances. For six errors or higher, the SER remains 100%, which is not unexpected, since it is harder to fix all the remaining errors in utterances that contain many
errors. Overall, the number of errors decreased in the utterances the re-scoring was applied