Improved convolutional recurrent network speech enhancement algorithm based on low complexity
Dun Liu, Changfei Sheng, Junfeng Wei · 2025
Aiming at the problems of existing speech enhancement methods in which the global feature modeling capability of speech signal is insufficient leading to unsatisfactory denoising effect, and the model computational complexity is high, a convolutional residual recurrent network speech enhancement method that integrates ECA and feature attention mechanism (ECA-CFCRRN) is proposed. The network encoder incorporates an efficient channel attention mechanism to synchronously improve the local spectral feature characterization capability and compress the parameter scale; the middle layer introduces residual structure to deepen the network and designs a feature attention mechanism to enhance the correlation between neighboring speech frames and global feature capture capability; and the joint time-frequency loss function is used to improve the quality of speech reconstruction. Experiments show that the method achieves better objective metrics on the publicly available dataset Voice Bank-DEMAND with 1.09 million parameters relative to other models, and the perceptual evaluation of speech quality (PESQ) is improved by 47.21% and 9.99dB improvement in SISNR compared to the original noisy speech; combined with the Perceptual Contrast Stretching, the PESQ is further improved by 7.93%, which verifies that the ECA-CFCRRN model is effective in improving speech quality and intelligibility.