U-Net网络中结合卷积与注意力机制的跳跃连接的单通道语音增强

Combining Convolution and Attention with Skip Connection in U-Net for Monaural Speech Enhancement

摘要: 在可懂度和感知质量方面，单通道语音增强技术从深度学习的成功中获得了巨大收益。传统的方法侧重于应用U-Net模型来预测带噪语音的纯净信号，这种模型的跳跃连接以及序列建模模块存在局限性。本研究提出了在U-Net网络中结合卷积与注意力机制的跳跃连接的单通道语音增强算法。一方面，基于卷积的跳跃连接(convolution skip)应用含有卷积门控机制的卷积模块来提取更重要的局部特征信息；另一方面，基于注意力机制的跳跃连接(attention skip)结合了ROPE位置编码与图卷积网络(GCN)，从而能够更好地提取上下文全局特征信息；除此之外，conformer-block模块应用了卷积门控单元(CGU)与多头增强的注意力机制，它能够更好的建模序列信息。在VoiceBank-DEMAND语音数据集上对提出的方法进行了验证，在噪声数据上获得了0.6975的语音感知质量评估(PESQ)提升、0.0124的语音短时客观可懂度(STOI)提升以及8.4324的分段信噪比(SSNR)提升。实验结果表明，与基线denoiser方法相比，提出来的方法更有优越性。

Abstract: In terms of intelligibility and perceptual quality, single-channel speech enhancement technology has benefited greatly from the success of deep learning. Traditional methods focus on applying the U-Net model to predict the clean signal from noisy speech, but this model’s skip connections and sequence modeling modules have limitations. This study proposes a single-channel speech enhancement algorithm that combines convolution and attention mechanisms in the skip connections of the U-Net network. On the one hand, the convolution skip connection applies a convolution module with a convolution gating mechanism to extract more important local feature information. On the other hand, the attention skip connection combines ROPE position encoding and graph convolutional networks (GCN) to better extract contextual global feature information. Besides, the conformer-block module applies a convolutional gated unit (CGU) and an enhanced multi-head attention mechanism, which can better model sequence information. The proposed method was validated on the VoiceBank-DEMAND speech dataset, achieving an improvement of 0.6975 in Perceptual Evaluation of Speech Quality (PESQ), 0.0124 in Short-Term Objective Intelligibility (STOI), and 8.4324 in Segmental Signal-to-Noise Ratio (SSNR) on noisy data. Experimental results show that the proposed method is superior to the baseline denoiser method.