Improved Conformer with Mamba-Like Linear Attention for Speech Recognition

The attention mechanism in the Conformer encoder is employed to model long-range dependencies in speech data. However, it requires global interactions across all input sequence elements to extract features, resulting in computational complexity that scales quadratically with the sequence length. In long-sequence speech recognition tasks, the high computational complexity considerably increases the model's training duration. To address this problem, we propose an end-to-end approach for speech recognition utilizing a gated linear attention mechanism. Firstly, SEFFN optimizes the channel-wise weight distribution, enhancing the efficient extraction and refinement of high-quality channel-specific features during the training process. The mechanism strengthens the model's capacity to prioritize critical channel-specific features, resulting in an improved capacity for high-quality channel feature modeling and representation. Secondly, a gated linear attention mechanism is employed to reduce the computational complexity to a linear relationship with the length of the speech sequence, thereby mitigating the computational overhead associated with the attention mechanism. Experimental results demonstrate that the character error rate of the model on the Chinese dataset Aishell-1 was reduced to 4.59%, representing a significant improvement compared to the baseline model, thereby validating the proposed method's effectiveness.