SPEECH SYNTHESIS AND RECOGNITION FOR A LOW-RESOURCE LANGUAGE Connecting TTS and ASR for mutual benefit

Abstract

Speech synthesis (text-to-speech, TTS) and speech recognition (automatic speech recognition, ASR) are the NLP technologies that are the least available for low-resource and indigenous languages. Lack of computational and data resources is the major obstacle when it comes to the development of linguistic tools for these languages. We present a framework that does not require enormous GPU and target data resources, as well as guarantees reasonably good results in performance for the end-product. In this work we perform dual connection between TTS and ASR models and make them learn from each other in a low-resource setup. This project, being the first open-source implementation of such a bidirectional algorithm, leverages the power of open-source projects for the benefit of indigenous languages. We release the first ever functioning ASR tool for the North Sámi language along with a competitive TTS technology, which fulfills the demand of the North Sámi community and globally contributes to the further development of AI tools for low-resource languages.