The NTNU Super Monster Team (SPMT) system for the Formosa Speech Recognition Challenge 2023 - Hakka ASR

This paper aims to record the progress of the NTNU Super Monster Team (SMPT) in the Formosa Speech Recognition Challenge 2023 (FSR-2023), which is the third event of the Formosa Speech in the Wild (FSW) project. The primary task was to recognize Hakka speech using a corpus of Hakka speakers in Taiwan. We present our participation results in Track 1: Taiwanese Hakka recommended characters speech recognition. Recently, the percentage of Hakka speakers in Taiwan is only about 5.5 percent of the total population, and is still decreasing year by year, which causes resistance in acquiring the corpus; due to the strong ethnic identity of the Hakka cultural group, it has a strong linguistic independence and exclusivity. In summary, the scarcity of Hakka pairedcorpus and the difficulty of learning other dialects for mutual benefit have undoubtedly aggravated the difficulty of the FSR-2023. In this study, we try to investigate the interleaving effects of various components by integrating data augmentation, self-supervised learning features, large-scale speech recognition models, and language models to improve the performance of Hakka speech recognition. This article aims to explore the impact of various modules on Hakka speech recognition performance and has ultimately achieved fruitful results. We hoped that this effort can contribute to the preservation of endangered languages in our country.