Hybrid Acceleration of CNN-based Speech Enhancement on Embedded Platforms

Speech enhancement is a crucial component in speech signal processing. Enhancement models based on neural networks greatly outperform traditional approaches at the cost of huge amounts of parameters and complex network structures, making it quite difficult to work on embedded platforms. To address this issue, this paper designs a convolution accelerating algorithm based on an autoencoder network called Redundant Convolutional-Encoder-Decoder (R-CED). Two methods have been designed to accelerate the computation, including a convolution structure of 'parallel data path + control logic + on-chip cache' and a parallelization acceleration strategy in convolution operation. The system is implemented on Xilinx Zynq 7020 platform to validate its effectiveness. Compared with commonly used Central Processing Unit (CPU) and Graphics Processing Unit (GPU) platforms, the processing delay of the accelerated enhancement algorithm is only 0.0016s, reduced by up to 99% while keeping the Perceptual Evaluation of Speech Quality (PESQ) score under 0.015, making it possible for real-time speech enhancement to be implemented on embedded platforms.