Efficient use of training data for sinhala speech recognition using active learning

Automatic Speech Recognition is an area which requires a large amount of training data. Collecting such quantities of data involves significant time and cost owing to the tedious nature of collecting speech recordings and manual nature of transcribing it. For a low resourced language such as Sinhala, collecting a sufficient data set is a major problem. To address this issue we used the Active Learning technique from the Machine Learning paradigm which is applied to many tasks such as information retrieval. Our experiment using a simple Sinhala speech corpus shows that through the use of Active Learning, the amount of utterances that need to be transcribed can be reduced by some 42% to achieve the same accuracy as using the whole data set without such a strategy. This suggests that Active Learning techniques can be successfully applied to make optimal use of scarce resources for speech recognition for new languages.