Pruning and Quantization Enhanced Densely Connected Neural Network for Efficient Acoustic Echo Cancellation

Acoustic echo cancellation (AEC) has gained prominence due to increasing commercial demand and diverse application scenarios. Apart from professional fields like audio processing, AEC finds more use in daily scenarios like video conferencing. However, attaining superior speech quality frequently involves intricate deep learning models and significant computational resources, thereby impeding its widespread adoption. In order to reduce model size and enhance inference speed while maintaining speech quality, in this paper, we propose a new network architecture that combines pruning and quantization with densely connected neural network called PQ-DCNN. The proposed PQ-DCNN model utilizes encoder unstructured pruning (EUP) modules and decoder structured pruning (DSP) modules to reduce model size and accelerate model inference. Static quantization (SQ) module targets convolution layers, optimizing model resources by converting less critical weights and activation values from high to low precision. Results demonstrate a 66.78% reduction in model size and a 26.46% decrease in inference time, achieving an efficient AEC model design.