English Speech Recognition Model Based on Improved Conformer

This paper proposes an end-to-end English speech recognition model based on improved Conformer, which solves the problem of audio-text misalignment and the defects of future information leakage in traditional speech recognition by fusing the occlusion multi-head attention mechanism and CTC joint training strategy. The model uses Conformer encoder as the core and combines relative position coding to capture long sequence context information. The masked multi-head attention is introduced into the decoder to ensure that the prediction only relies on generated tokens and avoids information errors. A multi-task learning framework is used to jointly optimize CTC loss and Conformer loss. Experiments based on English TTS dataset show that the improved model is significantly superior to the original Conformer in word error rate, and has stronger robustness, which can effectively suppress overfitting, and is suitable for actual English speech recognition scenarios.