Mathematical model and experimental evaluation of the stability of quantized DS-CNN for TinyML speech command recognition

The article compares 8- and 4-bit versions of DS-CNN for recognizing “yes” and “no” commands on a microcontroller. The proposed mathematical model relates the quantization error to the stability of the neural network and the limitations of its deployment.

The authors tested recognition accuracy under acoustic noise and gradient attacks in feature space. Of the tested options, the model trained taking into account 8-bit quantization received the highest integral score. The median latency on the Arduino Nano 33 BLE Sense was 41.53 ms.