Data Selection using Spoken Language Identification for Low-Resource and Zero-Resource Speech Recognition

Jianan Chen, Chenhui Chu, Sheng Li, Tatsuya Kawahara

Abstract: Large-scale pre-trained models have become common for Automatic Speech Recognition (ASR) tasks. They utilize large-scale, multilingual datasets to learn acoustic features and then are finetuned on downstream ASR tasks. However, their performance degrades when applied to low-resource and zero-resource languages lacking data. This paper introduces a new data selection method with Spoken Language Identification (SLI) models to bring non-target language speech data into training. With the help of phonetic labels as a universal intermediate representation to link low-resource languages to those with rich resources, we enhance the ASR system's performance on low-resource and even zero-resource languages. We conducted ASR experiments on Marathi, Assamese, and Panjabi with augmented non-target Hindi data in the CommonVoice corpora. The experimental results show that the proposed method can train ASR systems with little target language resource.