A Target-Speech-Feature-Aware Module for U-Net Based Speech Enhancement

<div class="section abstract"><div class="htmlview paragraph">Speech enhancement can extract clean speech from noise interference, enhancing its perceptual quality and intelligibility. This technology has significant applications in in-car intelligent voice interaction. However, the complex noise environment inside the vehicle, especially the human voice interference is very prominent, which brings great challenges to the vehicle speech interaction system. In this paper, we propose a speech enhancement method based on target speech features, which can better extract clean speech and improve the perceptual quality and intelligibility of enhanced speech in the environment of human noise interference. To this end, we propose a design method for the middle layer of the U-Net architecture based on Long Short-Term Memory (LSTM), which can automatically extract the target speech features that are highly distinguishable from the noise signal and human voice interference features in noisy speech, and realize the targeted extraction of clean speech. Then, in order to achieve deep fusion between the target speech features and the model, we design a multi-scale deep fusion skip connection method, so that when the effective information flows from the encoder to the decoder, the features with large correlation with the target speech are effectively screened through the weight coefficient of attention. Finally, in order to verify the effectiveness of the proposed module, experiments were carried out on the Voicebank+Demand speech dataset. The results show that the proposed method has strong robustness in the environment with human voice interference. It outperforms other algorithms on metrics such as PESQ, STOI, CSIG, CBAK, COVL, offering cleaner speech with higher perceptual quality and intelligibility. This makes it particularly promising for applications in scenarios with significant human voice interference, such as in-car environments.</div></div>