Simple Data Augmented Transformer End-To-End Tibetan Speech Recognition

The Tibetan language belongs to the Tibetan branch of the Sino-Tibetan language family, and which is the common language of the Tibetan people in China, mainly containing the Ando, U-Tsang and Kham dialects. The Tibetan language has its own phonetic system, grammatical structure, rich vocabulary, and perfect expressive ability. This paper builds a corpus of U-Tsang Tibetan and designs corresponding textual annotations. Based on this corpus, a Transformer network-based method for Tibetan speech recognition is proposed and a Transformer end-to-end speech recognition model for Tibetan is trained. We compared the Transformer-based approach to Tibetan speech recognition with the convolutional neural network (CNN)-based model and introduced a data augmentation algorithm SpecAugment in training. The results show that under the use of 36000 Tibetan utterances corpora with a total length of 53h, transformer model and CNN model respectively obtain 29.3% and 32.6% of WER. After using the data augmentation algorithm SpecAugment, we got 25.8% and 28.1% in Transformer model and CNN model respectively.