SlimNet: A Lightweight Attentive Network for Speech-Music-Noise Classification and Voice Activity Detection

In this work, we present SlimNet, a hybrid architecture that incorporates a long short-term memory with an attention mechanism for speech-music-noise classification and speech vs non-speech categorization. Deep neural network approaches for these tasks have achieved remarkable success. However, deep models are computationally and energy intensive. When compared to state-of-the-art models, SlimNet uses up to 7 times fewer parameters, with up to 5 times faster inference while achieving comparable performance. The small footprint of SlimNet makes it a simple yet powerful candidate for devices with limited computational resources, and for edge computing devices in particular.