The impact of quantization of TinyML models on the stability of audio systems of the personal Internet of things

The report examines the impact of different quantization modes of machine learning models on the stability of Personal Internet of Things (PIoT) audio systems operating in conditions of limited computing resources. The target devices are PIoT devices equipped with microphones, such as smart headphones and speakers with voice assistants that use local TinyML models to recognize audio commands. An experimental analysis of the stability of models and their variations using quantization under the influence of adversarial audio disturbances was carried out.