Comparative Study of Quantized CNN Inference on ARM and RISC-V Microcontrollers

This paper presents a comparative study of a lightweight Convolutional Neural Network's (CNN) performance, for speech recognition, deployed in three microcontroller platforms with resource constraints. The devices used were the STM32F411CEU6 (ARM Cortex-M4), the Seeed Studio XIAO ESP32-C6 (dual-core RISC-V) and the RP2350 (dual-core Cortex-M33 and dual-core RISC-V). This study evaluates model time inference and energy consumption, with both quantized (Int8) and non-quantized (Float32) variants. To the best of the author's knowledge, this is the first implementation of an iterative spectrogram computation technique specifically optimized for low-memory microcontroller environments performing real-time inference. Additionally, the Pico TensorFlow Lite port was extended to support inference on the RISC-V cores of the Pico 2 W (RP2350), enabling performance comparisons between ARM and RISC-V architectures on a shared platform. Results reveal the trade-offs between ARM and RISC-V based microcontrollers regarding speed, power consumption, implementation complexity and overall performance, offering practical insights for the deployment of speech based AI on the edge.