A New U-Net Speech Enhancement Framework Based on Correlation Characteristics of Speech

As a key component of in-vehicle intelligent voice technology, speech enhancement can extract clean speech signals contaminated by environmental noise to improve the perceptual quality and intelligibility of speech. It has extensive applications in the field of intelligent car cabins. Although some end-to-end speech enhancement methods based on time domain have been proposed, there is often limited consideration given to designing model architectures based on the characteristics of the speech signal. In this paper, we propose a new U-Net based speech enhancement framework that utilizes the temporal correlation of speech signals to reconstruct higher-quality and more intelligible clean speech. Firstly, to address the issue of inadequate extraction of multi-scale correlation features from speech signals during feature extraction and reconstruction, a novel dense connection multi-scale feature extraction module based on gated dilated convolution is devised to enhance temporal receptive length and extract diverse scale features effectively. Secondly, in order to tackle the problem of feature loss and harmonic distortion during sampling, a sophisticated pooling-reconstruction fine-grained sampling method based on feature map recombination is proposed. This method aims to minimize information loss during down-sampling while simultaneously enhancing the clarity of reconstructed waveforms during up-sampling. Lastly, leveraging the aforementioned pooling-reconstruction sampling method, we propose a deep supervision approach for multi-scale feature. This approach effective supervision of perception characteristics across different frequency ranges. In order to validate the effectiveness of the proposed framework, experiments were conducted on the Voicebank+Demand dataset. The results show that compared to other advanced algorithms, the proposed model significantly improves metrics such as PESQ, STOI, CSIG, CBAK, and COVL. Even in low SNR environments, the enhanced speech signals exhibit noticeable improvements in quality and intelligibility. This is beneficial for subsequent automotive voice applications.