Exploring Model-Level Transfer Learning to Improve the Recognition of Sinhala Speech

This chapter proposes a transfer-learning-based approach to implementing an ASR system for Sinhala using Mozilla’s DeepSpeech. In this chapter, the focus is on identifying the most appropriate target model for Sinhala speech prediction with the support of a language model and data augmentation. The dataset used for this chapter contained 2000 unique sentences from 123 speakers. Those sentences were recorded with the support of Praat and RedStart tools. All the experiments were conducted with an external 4-gram language model for Sinhala using the KenLM toolkit. Without applying any augmentation, our transfer-learned model showed better results, achieving 22.92% in WER and 8.84% in CER. Then it showed a further improvement on prediction by achieving 17.19% in WER and 5.9% in CER using data augmentations with the probabilities of reverb at 0.1 and overlay at 0.3 with scorer optimizations.