MCodec: A Codec Architecture integrated Mamba for Speech Enhancement

Recently, state space models have made significant progress in modeling sequential tasks, particularly the SSM-based Mamba network. Leveraging the efficient modeling of sequential data by the Mamba, we propose an end to end convolution-based codec architecture called MCodec that incorporates Mamba for speech enhancement tasks. The encoder of codec directly models the raw audio sampling points, performing downsampling and feature extraction akin to Mel-frequency spectra. Meanwhile, the decoder part, symmetrically mirroring the encoder, is utilized to reconstruct the waveform from the feature space, reducing reliance on manual feature extraction. Additionally, a bidirectional Mamba(Bi-Mamba) is introduced between these two components to extract deeper dependencies for speech reconstruction, effectively denoising the speech in the feature space. Experimental results demonstrates that the proposed model MCodec achieves promising improvements on the VoiceBank+DEMAND dataset.