Quantized ID-CNN for a Low-power PDM-to-PCM Conversion in TinyML KWS Applications

This paper proposes a novel low-power HW accelerator for audio PDM-to-PCM conversion based on artificial neural network. The system processes samples from a digital MEMS microphone and converts them in PCM format by using a 1-Dimensional Convolutional Neural Network (1D-CNN). The model has been quantized to reduce the computational complexity while preserving its Signal-to-Noise Ratio (SNR) and the HW accelerator has been designed to minimize the physical resources. The SNR achieved is 41.56 dB while the prototyping of the design on a Xilinx Artix-7 FPGA shows a dynamic power consumption of 1 mW and a utilization of 606 LUTs and 410 FFs. These results enable the proposed system to be the first step of a tiny low-power end-to-end neural network-based Keyword Spotting (KWS) system.