End-to-End Speech Recognition of Central Miao Language Based on Wav2Vec 2.0

Miao, as one of China's minority languages, has a complex phonetic system and limited corpus resources, which makes existing speech recognition methods difficult to meet practical demands. This paper introduces the application of the Wav2Vec 2.0 framework for the first time to perform end-to-end speech recognition for the Miao language. We pre-trained the wav2vec 2.0 model on a large amount of unannotated Miao speech data, and then fine-tuned it with a small amount of annotated data. To advance research in Miao speech recognition, this paper also constructs a new 20-hour Miao speech dataset, DHD, and conducts experiments on it. The experimental results show that, compared to traditional end to end speech recognition methods, the Wav2Vec 2.0-based model achieves a WER of 31.3 % on the Miao speech recognition task, demonstrating a significant improvement. Our method not only provides a new solution for Miao speech recognition but also offers a feasible framework for speech processing of other resource-scarce languages.