Multi Source Speech Enhancement for Low-Power Micro-Controller devices

In the latest years, speech-centric sensors, that combine audio sensing capabilities and local processing, are getting popular in consumer applications. Filtering out unwanted noises or extracting relevant features in real time is extremely challenging because of the severe energy constraints and the limited memory and computation capacities of typically used digital processing platforms. To this aim, this work investigates novel neural-based methods for speech processing. In particular, the technique proposed belongs to the Speech Enhancement field and leverages microphone array processing. In this context, Beamforming is the reference technique and state-of-the-art models obtain effective results using neural networks. Thus we propose a neural beamforming model, tailored for lightweight implementation on low-power microcontrollers (MCU). In this work, we first show the performances of the developed model against a synthetically generated dataset and then we study its deployment as a streaming application on a real MCU.