Comparative Analysis of Branchformer and EBranchformer for Tibetan Speech Recognition

End-to-end speech recognition has advanced in high-resource languages but still faces limitations in low-resource languages like Tibetan due to scarce speech data. This paper preprocesses the Lhasa Tibetan dataset with cleaning and deduplication, and uses data augmentation to expand the training set for an effective speech recognition system.This paper constructs an end-to-end speech recognition model for the Tibetan language based on the WeNet framework, and systematically evaluates the recognition performance of Branchformer and its enhanced variant, E-Branchformer. During the decoding phase, four strategies were employed:attention, CTC greedy search, prefix beam search, and attention rescoring. Experimental results demonstrate that E-Branchformer outperforms Branchformer across all decoding methods, achieving relative reductions in word error rate of 1.08%, 1.23%, 1.07%, and 1.05%, respectively. These findings validate that E-Branchformer exhibits stronger feature modeling capabilities and higher recognition accuracy in low-resource language recognition tasks.