VAD to the Bone: Ultra-Tiny Speech Activity Detection for Edge Deployment

TL;DR - kiloVAD is a 2.1K-parameter, CNN-only voice activity detector designed for causal embedded inference. It achieves 0.850 AUC on AVA-Speech while using standard, deployment-friendly components.

• Uses standard Mel features and avoids recurrent layers, learnable filterbanks, and non-causal post-processing.
• Combines per-layer structured pruning with self-distillation.
• Angle-based quantization-aware training improves results by 1–4% over standard QAT.
• Operates with 200 ms of context under causal, per-frame evaluation.