The Analysis of End-to-End Tibetan Speech Recognition Based on Speaker Representation and Adaptive Technology

Tibetan speech recognition currently suffers from low accuracy and a severe lack of training data, a challenge that becomes particularly pronounced when there is a mismatch between testing and training conditions. To enhance the performance of Tibetan speech recognition, this study proposes two speaker adaptation strategies: feature-domain adaptation and model-domain adaptation. First, leveraging a Tibetan speech dataset, we extracted three distinct speaker representations (i-vector, x-vector, and ECAPA-TDNN) and conducted feature-domain adaptation experiments using feature addition (+) and feature concatenation (⊕) techniques. The experimental results demonstrate that feature-domain adaptation effectively improves recognition performance, with the ECAPA-TDNN embedding yielding the most significant gains. Subsequently, a model-domain adaptation approach was implemented by integrating Learning Hidden Unit Contributions (LHUC) with x-vector and ECAPA-TDNN. This method further reduced the Character Error Rate (CER), outperforming the feature-domain adaptation methods. Building on these findings, we combined feature-domain and model-domain adaptation to capitalize on their synergistic advantages. The results show that the proposed hybrid method based on ECAPA-TDNN achieved a relative CER reduction of 47.60% (from 14.18% to 7.43%) compared to the baseline system. This research demonstrates that the integration of feature and model-domain adaptation effectively mitigates the data mismatch problem, significantly improving the accuracy of Tibetan speech recognition.