Transfer learning from English to Slovak in speech recognition applications

The speed of research can be slowed down by limited access to SOTA computational resources. This paper compares two different speech recognition approaches, training from scratch and transfer learning. It is focused on the transfer learning technique known as fine-tuning in cross-language adaptation speech recognition tasks. The main idea of this paper is to prove the feasibility of adapting the foreign speech recognition model to the Slovak ASR model. We analyze the benefits of BPE tokenization in speech recognition, which we have not previously tested, and a particular pre-trained model in terms of training speed and model quality. In practice, the publicly available pre-trained English Librispeech100 model is used to create a transferred English-to-Slovak model on TUKE-BNews-SK corpus subset. In addition, the paper provides details on the ESPnet framework workflows involved in the creation of the proposed end-to-end fine-tuned model.