Semi-supervised training for lecture transcription in resource-scarce environments

Abstract

We present a study where standard semi-supervised training methods are applied in a resource-scarce environment to build lecture transcription systems. Experiments are conducted on two different corpora which one can expect to be available in resource-scarce environments. These include 1) speaker- and domain-specific data where a single South African English lecturer presents the “Operating Systems” course, and 2) Afrikaans speaker-independent and domain non-specific data collected from science and law courses. Different amounts of acoustic and language model data are used for training the respective models. We find that lecture transcription systems in resource-scarce environments can benefit substantially from semi-supervised training methods. We also describe a small, new corpus of spoken lectures which is freely available in the public domain.