A Neural Network Ensemble Diagnostic Evaluation of TIMIT Speech Corpus

TIMIT is a significant dataset in the field of Automatic Speech Recognition and has been used to implement and evaluate many important developments in the field. Although now 30 years old, TIMIT remains an important corpus in the field of human language technologies. One reason for this is that it provides speech audio with broad phonetic coverage, transcriptions and high quality detailed phone-level annotations. This thesis concerns some interesting classifications produced by an ensemble of articulatory feature recogniser neural network models trained and evaluated using TIMIT. Examination of those segments unanimously misidentified by the ensemble finds that due to the articulatory and mechanical constraints of human speech a small part of the dataset does not, and possibly can not, align with the phonetic annotations in the corpus. As experiments publish results returning accuracy rates in excess of 90%, an intrinsic consequence of the unavoidable properties of a small part of the dataset overall can contribute a significant proportion of the evaluated error rate. This should therefore be considered by researchers working with this valuable resource.