Accent Classification Using Spiking Neural Networks

Authors: Cosmin Moarcas

Presentation type: Poster at SNUFA 2024 online workshop (5-6 Nov 2024)

Abstract

In recent years, the field of artificial intelligence has experienced significant growth. With the integration of AI systems across various domains, the energy consumption required for training these systems has also increased considerably. To address the issue of high energy consumption, companies such as Intel and CyberSwarm from Romania have developed neuromorphic chips designed to enhance the efficiency of conventional systems through an innovative architecture inspired by the functioning principles of the brain. These chips facilitate the training of neural networks that are different from traditional artificial ones, called spiking neural networks. This paper aims to develop and evaluate spiking neural networks for accent classification in the English language. By comparing deep neural networks with spiking ones, we demonstrated that both achieve similar performance levels, but the latter offer the advantage of reduced energy consumption during training. Accent classification models can substantially improve the performance of voice recognition systems like Alexa and Siri. By detecting accents, specialized systems can be developed for each accent type, making these technologies more accessible to individuals with pronounced accents.