Hybrid Neural Networks for On-device Directional Hearing

On-device directional hearing requires audio source separation from a givendirection while achieving stringent human-imperceptible latency requirements.While neural nets can achieve significantly better performance than traditionalbeamformers, all existing models fall short of supporting low-latency causalinference on computationally-constrained wearables. We present DeepBeam, ahybrid model that combines traditional beamformers with a custom lightweightneural net. The former reduces the computational burden of the latter and alsoimproves its generalizability, while the latter is designed to further reducethe memory and computational overhead to enable real-time and low-latencyoperations. Our evaluation shows comparable performance to state-of-the-artcausal inference models on synthetic data while achieving a 5x reduction ofmodel size, 4x reduction of computation per second, 5x reduction in processingtime and generalizing better to real hardware data. Further, our real-timehybrid model runs in 8 ms on mobile CPUs designed for low-power wearabledevices and achieves an end-to-end latency of 17.5 ms.