Speech Signal Enhancement Using Progressive Learning and Dense Connected LSTM Networks

Abstract

Aiming at the deficiencies of traditional speech signal enhancement models in dealing with long-term dependencies and noise filtering, an application speech signal enhancement model based on progressive learning and dense connection strategies is proposed. This method takes the long short-term memory network structure as the core and realizes the gradual enhancement of noisy speech through layer-by-layer learning and processing. The experimental results showed that this model exhibited excellent enhancement performance in different signal-to-noise ratio environments. In a -5dB signal-to-noise ratio environment, the short-term objective clarity of the research method reached 0.930, which was 4.1% higher than that of delayed neural networks. Moreover, under the 10dB condition, the short-term objective clarity score further increased to 0.957. The distortion signal ratio of the source signal has increased from 2.31 at -5dB to 14.81 at 10dB, indicating the model's ability in noise suppression and signal reconstruction. The assessment score of speech quality perception increased from 1.86 at -5dB to 3.13 at 10dB, and the word error rate decreased to 27.31%, which was 2.47% lower than that of the classical long short-term memory network. The research results show that the proposed model has strong robustness and a good speech enhancement effect when dealing with speech signals with a low signal-to-noise ratio, providing a new solution for the field of applied language processing.