【摘要】 针对人工耳蜗在噪声环境下言语感知效果差，以及现有算法降噪能力不足的问题，本研究提出了一种改进的Wave-U-Net模型。通过采取轻量化卷积，引入注意力机制，改进损失函数，优化数据集结构，以提高人工耳蜗的降噪效果。使用短时客观可懂度(short-time objective intelligibility, STOI)、语音质量评估(perceptual evaluation of speech quality, PESQ)、浮点运算次数(floating point operations per second, FLOPs)和参数量(Params)对模型的降噪效果和复杂度进行了评估，分别达到0.81、2.75,0.83 G,1.04 M。实验结果表明，本研究算法在符合人工耳蜗产品规范的基础上，实现了明显的降噪效果，提高了人工耳蜗使用者在复杂噪声环境中的语音感知效果。本研究方法为人工耳蜗算法的改进提供了新的可能，可为听力受损患者提供更好的听觉感受。更多还原

【Abstract】 Aim at poor speech perception in noisy environments by cochlear implant users and the inadequacy of existing noise reduction algorithms, we proposed an improved Wave-U-Net model. By adopting lightweight convolution, introducing attention mechanism, improving loss function, and optimizing dataset structure, the noise reduction effect of cochlear implants was enhanced. Using short-time objective intelligibility(STOI), perceptual evaluation of speech quality(PESQ), floating point operations per second(FLOPs), and Params to evaluate the noise reduction effect and complexity of the model, which reached 0.81, 2.75, 0.83 G, and 1.04 M, respectively. The experimental results showed that our algorithm achieved obvious noise reduction effect while conformed to the specifications of cochlear implant products, improved the speech perception effect of cochlear implant users in complex noise environments. The results provide new possibilities for the improvement of cochlear implant algorithms and can offer better auditory experience for patients with hearing impairment.更多还原