Designing and evaluating methods for the detection of mistakes in Sprekend Nederland data

Abstract

Due to a constant increase in audio and video information, the demand for methods that automatically align and filter this information keeps growing. Towards that end, this thesis aims to fulfill two objectives. First, to show the accuracy of the Spraaklab aligner and use it to align the Sprekend Nederland corpus, so that it can be used in further research. Second, to design and evaluate a method for automatically detecting user made mistakes in read speech using Spraaklab’s alignment process. The Corpus Gesproken Nederlands is used to develop alignment accuracy and mistake detection benchmarks. Spraaklab is shown to be accurate in aligning Sprekend Nederland and the read speech data is aligned. A mistake detection method using word recognition scores is developed, and shown to be effective on the Corpus Gesproken Nederlands. Due to score calibration problems it can not be shown to be effective on Sprekend Nederland, but the results indicate that further research could be able to show it, given more manually verified Sprekend Nederland alignments to establish better thresholds.