PROJECT TITLE :

Analysing Acoustic Model Changes For Active Learning In Automatic Speech Recognition - 2017

ABSTRACT:

In active learning for Automatic Speech Recognition (ASR), a portion of information is automatically selected for manual transcription. The objective is to boost ASR performance with retrained acoustic models. The commonplace approaches are based on confidence of individual sentences. During this study, we investigate an alternate view on transcript label quality, in which Gaussian Supervector Distance (GSD) is used as a criterion for data choice. GSD may be a metric which quantifies how the model was changed during its adaptation. By using an automatic speech recognition transcript derived from an out-of-domain acoustic model, unsupervised adaptation was conducted and GSD was computed. The custom-made model is then applied to an audio book transcription task. It is found that GSD offer hints for predicting information transcription quality. A preliminary try in active learning proves the effectiveness of GSD choice criterion over random selection, shedding light-weight on its prospective use.