NAS-VAD: Neural Architecture Search for Voice Activity Detection

The need for automatic design of deep neural networks has led to the emergence of neural architecture search (NAS), which has generated models outperforming manually-designed models. However, most existing NAS frameworks are designed for image processing tasks, and lack structures and operations effective for voice activity detection (VAD) tasks. To discover improved VAD models through automatic design, we present the first work that proposes a NAS framework optimized for the VAD task. The proposed NAS-VAD framework expands the existing search space with the attention mechanism while incorporating the compact macro-architecture with fewer cells. The experimental results show that the models discovered by NAS-VAD outperform the existing manually-designed VAD models in various synthetic and real-world datasets. Our code and models are available at https://github.com/daniel03c1/NAS_VAD.