BULLET: A Fast and Lightweight End-to-End Neural Network for Joint Acoustic Echo Cancellation and Noise Suppression
Song Chen, Yinlong Xu · 2024
Acoustic echo cancellation (AEC) and noise suppression (NS) are critical components in modern full-duplex real-time communication systems. Traditionally, these tasks are handled separately using large network, posing challenges for simultaneous deployment on mobile devices due to high computational demands. In this paper, we present Bullet, a lightweight end-to-end model designed to address both AEC and NS tasks. We leverage the shared features between AEC and NS, integrating both tasks into a single network to reduce computational complexity. Band Merge (BM) is used to further reduce complexity by compressing less perceptible high-frequency information, achieving a 46.9% reduction in computational load without compromising performance. To further improve the model's performance, our model incorporates perceptual loss and a Complex Ratio Masking (CRM) module, enhancing subjective quality. We design extensive data augmentation techniques to improve generalization across a variety of real-world scenarios. Experimental results demonstrate that Bullet with only 0.19M parameters achieves real-time inference on a single thread of an Intel Core i7-12700H @2.3GHz CPU, with a real-time factor (RTF) of 0.025, while maintaining competitive performance. Moreover, Bullet has been successfully deployed on actual mobile devices such as laptops and smartphones with minimal power consumption.