Single-Channel Speech Enhancement with Conformer
Haining Wang, Ying Gao, Qing Qi Zhao, Shifeng Ou · 2023
The transformer model is renowned for performing well in the field of natural language processing, thanks to its capability of extracting global features. Initially, conformer model was proposed for separating continuous speech. Inheriting the transformer model's ability to extract global features, the conformer model is equipped with a convolutional layer structure that can extract local features with more efficacy. This study aims to introduce a waveform-level end-to-end single-channel speech enhancement approach based on conformer. The proposed approach is made up of three distinct components: the downsampling layer, the conformer layer, and the upsampling layer. The conformer layer is composed of multiple conformer blocks that leverage residual connections in both parallel and series. The ablation experiments demonstrate that the conformer layer with series-parallel connections significantly improves speech enhancement performance. The experimental results, employing the VoiceBank-DEMAND corpus dataset, indicate that the proposed approach outperforms other competing techniques in terms of speech enhancement performance.