Real-Time Low Power Audio Distortion Circuit Modeling: a TinyML Deep Learning Approach

This work proposes a combined approach using deep learning and standard signal conditioning to model analog audio distortion circuits in real-time. The proposed model targets low latency and low power resource-constrained processors, and it is based on a quantized neural network. Pre- and post-processing consisting in filtering and dithering has been applied to improve the performance of the proposed optimized model. The model has been compared with a real-time state-of-art model WaveNet3 running on a modern desktop computer, which we used as performance reference. Our proposed model achieves a parameter size reduction of 8.4x and a neural network multiply-accumulation operations reduction of 5.1x with respect to the WaveNet3, while maintaining comparable performance. The model has been optimized and deployed on the novel MAX78000 microcontroller which features an on-board convolutional neural network hardware accelerator. Experimental results show real-time operation with a total latency of less than 10 ms and a power consumption as low as 91.8 mW in active mode, making the system suited for live music performances. Experimental results also demonstrate the capability of the hardware accelerator of the MAX78000 to reach 53.2 multiply-accumulation operations per cycle.