Year: 2025
Type: article
Abstract: This paper proposes EPyNet, a deep learning architecture designed for energy reduced audio emotion recognition.In the domain of audio based emotion recognition, where discerning emotional cues from audio input is crucial, the inte- gration of artificial intelligence techniques has sparked a transformative shift in accuracy and performance.Deep learn- ing, renowned for its ability to decipher intricate patterns, spearheads this evolution. However, the energy efficiency of deep learning models, particularly in resource-constrained environments, remains a pressing concern. Convolutional operations serve as the cornerstone of deep learning systems. However, their extensive computational demands leading to energy-inefficient computations render them as not ideal for deployment in scenarios with limited resources. Ad- dressing these challenges, researchers came up with one-dimensional convolutional neural network (1D CNN) array convolutions, offering an alternative to traditional two-dimensional CNNs, with reduced resource requirements. How- ever, this array-based operation reduced the resource requirement, but the energy-consumption impact was not studied. To bridge this gap, we introduce EPyNet, a deep learning architecture crafted for energy efficiency with a particular emphasis on neuron reduction. Focusing on the task of audio emotion recognition, We evaluate EPyNet on five pub- lic audio corpora—RAVDESS, TESS, EMO DB, CREMA D, and SAVEE.We propose three versions of EPyNet, a lightweight neural network designed for efficient emotion recognition, each optimized for different trade-offs between accuracy and energy efficiency. Experimental results demonstrated that the 0.06M EPyNet reduced energy consumed by 76.5% while improving accuracy by 5% on RAVDESS, 25% on TESS, and 9.75% on SAVEE. The 0.2M and 0.9M models reduced energy consumed by 64.9% and 70.3%, respectively. Additionally, we compared our Proposed 0.06M system with the MobileNet models on the CIFAR-10 dataset and achieved significant improvements. The proposed system reduces energy by 86.2% and memory by 95.7% compared to MobileNet, with a slightly lower accuracy of 0.8%. Compared to MobileNetV2, it improves accuracy by 99.2% and reduces memory by 93.8%. When compared to MobileNetV3, it achieves 57.2% energy reduction, 85.1% memory reduction, and a 24.9% accuracy improvement. We further test the scalability and robustness of the proposed solution on different data dimensions and frameworks.
Authors: Jiby Mariya Jose, Jeeva Jose
Language: English
Topic: Emotion and Mood Recognition
Subfield: Experimental and Cognitive Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: diamond
OpenAlex ID: W4411523820