Evaluating the energy consumption of LPC-based and VAE-GAN timbre transfer models in Music AI

Sammanfattning

Energy efficiency in Music AI is becoming an increasingly important concern, especially as deep learning models continue to grow in size and complexity. Timbre transfer models, which transform one instrument’s sound to another, hold great potential in advanced music synthesis as well as growing energy demands. This thesis investigates two timbre transfer approaches: an LPC- based model representing classic signal processing methods, and a VAE- GAN model leveraging deep learning techniques. The aim is to quantify and compare their energy usage during both training and inference, raising awareness of energy-efficient practices in this domain. This study employed a smart plug to monitor energy consumption for timbre transfer models, capturing power usage during both CPU-based parallel processing for the LPC model and GPU-intensive training and inference of the VAE-GAN model. As the first investigation into smart plug-based energy measurement within the Music AI field, this research offers a novel perspective on energy assessment for music synthesis technologies. However, while the smart plug method provides an accessible and affordable solution, it has notable limitations such as ignorance of details during data processing and limited temporal resolution. Further development of high-resolution hardware and software tools would facilitate more precise measurement and analysis. Results indicate a substantial difference in energy consumption between the two models, with the LPC-based model showing higher efficiency in limited-data contexts and the VAE-GAN model excelling in complex and large-scale timbre transformations at the cost of increased energy usage. These findings underscore the importance of model design and dataset considerations for developing less resource intensive Music AI applications, promoting a balance that supports both artistic creativity and environmental responsibility.