Hierarchical Softmax for End-to-End Low-resource Multilingual Speech Recognition
Hierarchical Softmax for End-to-End Low-resource Multilingual Speech Recognition

概要

Low-resource speech recognition has been long-suffering from insufficient training data. In this paper, we propose an approach that allows neighboring languages to help improve low-resource scenario performance. We assume similar units in neighbor languages share similar term frequency and form a Huffman tree to perform multilingual hierarchical Softmax decoding. The hierarchical structure can allow similar tokens to share knowledge across languages, thus benefiting low-resource training. Experimental results show that our method can improve the accuracy and efficiency of low-resource speech recognition.

産業界への展開例・適用分野

音声認識

研究者

氏名	コース	研究室	役職/学年
Zhengdong Yang	知能情報学コース	言語メディア研究室	博士1回生