Deep Learning Based End-to-End Tibetan Speech Recognition Research

With the rapid development of information technology, fast and accurate recognition of speech signals by machines has become the basis of human-computer interaction. As one of the important technologies of human-computer interaction, speech recognition technology is gradually applied to minority languages. Tibetan is the main language of the Tibetan people in their daily life, and belongs to the Tibetan branch of the TibetoBurman language family of the Sino-Tibetan language family. The prevalent Tibetan dialect in Tibet is the Wei-Zang dialect represented by Lhasa. As one of the minority languages in China, the research on phonological recognition of Tibetan is relatively late and the scale of the data is small, so the traditional end-to-end phonological recognition model does not perform satisfactorily in the task of phonological recognition of Tibetan. In this paper, we propose an end-to-end Tibetan speech recognition model based on Conformer encoder and CTC/Attention hybrid architecture. The proposed model is validated with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 0}$</tex> hours of Tibetan speech data volume, and the experimental results show that the word error rate is relatively reduced by 12.43 % and 6.16 % on the test set compared to the Conformer and Conformer-CTC models, respectively, and the final word error rate of the model is <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2 2. 1 3 \%}$</tex>.