In this paper, speech enhancement is carried out using a modified nonlocal means adaptive filter, and an FPGA prototype is proposed for its implementation. Nonlocal means (NLM) adaptive filtering has been a widely used approach for denoising image and electrocardiogram signals. The temporal and spectral characteristics of the speech signal vary with time. Consequently, NLM filtering is ineffective in enhancing the noisy speech signal. In this study, the weight values estimated during NLM filtering are nonuniformly smoothed through energy-dependent scaling to address this issue. The proposed modification in NLM adaptive filtering significantly improved speech enhancement performance compared to other classic NLM filtering and enhancement methods. The proposed modified NLM technique provides better efficacy for objective quality measures by observing the simulation results presented on the TIMIT database. This study is performed on signals corrupted with white, babble, and factory noises at different SNR levels taken from the Noisex-92 database. The hardware architecture of the proposed method is designed and verified by implementing it on a Zynq-7000 (xc7z045ffg900-2) FPGA using the Xilinx system generator.

在本文中，语音增强是使用改进的非局部均值自适应滤波器进行的，并提出了一个 FPGA 原型来实现它。非局部均值 (NLM) 自适应滤波已成为一种广泛用于图像和心电图信号去噪的方法。语音信号的时间和频谱特性随时间变化。因此，NLM 滤波在增强带噪语音信号方面是无效的。在这项研究中，在 NLM 滤波期间估计的权重值通过依赖能量的缩放来非均匀平滑以解决这个问题。与其他经典的 NLM 滤波和增强方法相比，在 NLM 自适应滤波中提出的修改显着提高了语音增强性能。通过观察 TIMIT 数据库上显示的模拟结果，所提出的改进 NLM 技术为客观质量测量提供了更好的效果。这项研究是对从 Noisex-92 数据库中获取的不同 SNR 级别的白噪声、杂音和工厂噪声破坏的信号进行的。通过使用赛灵思系统生成器在 Zynq-7000 (xc7z045ffg900-2) FPGA 上实现，设计并验证了所提出方法的硬件架构。