Selection of Supplementary Acoustic Data for Meta-Learning in Under-Resourced Speech Recognition
I-Ting Hsieh, Chung‐Hsien Wu, Zhe-Hong Zhao · 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) · 2022
Automatic speech recognition (ASR) for under-resourced languages has been a challenging task during the past decade. In this paper, regarding Taiwanese as the under resourced language, the speech data of the high-resourced languages which have most phonemes in common with Taiwanese are selected as the supplementary resources for meta-training the acoustic models for Taiwanese ASR. Mandarin, English, Japanese, Cantonese and Thai as the high-resourced languages are selected as the supplementary languages based on the designed selection criteria. Model-agnostic meta-learning (MAML) is then used as the meta-training strategy. For evaluation, when 4000 utterances were selected from each supplementary language, we obtained the WER of 20.89% and the SER of 8.86% for Taiwanese ASR. The results were better than the baseline model (26.18% and 13.99%) using only the Taiwanese corpus and traditional method.