Formosa Speech Recognition Challenge 2020 and Taiwanese Across Taiwan Corpus

摘要

Taiwanese (a.k.a. Taiwanese Hokkien, Hoklo, Taigi, Southern Min or Min-Nan) is an endangered language, because the domination of Mandarin, the number of Taiwanese speakers continues to drop, especially among the youth generations. In addressing this problem, a Taiwanese speech-enabled human-computer interface for supporting people's daily life is essential. Therefore, a Formosa Speech in the Wild (FSW) project was established to collect a large-scale Taiwanese speech across Taiwan (TAT) corpus to boost the development of Taiwanese speech recognition (TSR). A Formosa Speech Recognition Challenge 2020 (FSR-2020) was also hosted to promote the corpus as well as to evaluate the performance of state-of-the-art TSR systems. This paper briefly introduces TAT corpus and FSR-2020 challenge, presents the provided data profile, evaluation plan and reports experimental baseline results. A subset of TAT corpus, TAT-Vol1, is given away for free for all participants (non-commercial license), and its corresponding Kaldi baseline recipes have been published online. Experimental results have showed that the combination of TAT corpus and the baseline recipes is a good resource pack for TSR research and development.