kiloVAD sets new edge record with 2.1k-parameter voice activity detection
A 2,100-parameter model beats larger VADs using only standard Mel features and CNNs.
Voice activity detection (VAD) is the gatekeeper for always-on speech systems, but existing compact models often rely on untrustworthy components like learnable filterbanks or recurrent layers. In a paper accepted at INTERSPEECH 2026, a team of researchers introduces kiloVAD, an ultra-tiny VAD designed for embedded inference. The model uses only standard Mel features and CNN layers, with tunable context and spectral parameters, ensuring broad hardware support. They also introduce per-layer structured pruning with self-distillation and an angle-based quantization-aware training (QAT) method that outperforms standard QAT by 1–4%.
Evaluated per-frame under causal conditions on the AVA-Speech benchmark, kiloVAD achieves 0.850 AUC with just 2.1k parameters and 200ms of context—a new state of the art for casual, deployment-ready VAD. This efficiency means the model can run on microcontrollers and edge devices without sacrificing accuracy, enabling battery-powered devices to listen continuously for wake words or triggers. The work highlights how clever training techniques and architectural minimalism can beat larger models in resource-limited settings.
- Uses only standard Mel features and CNN layers, avoiding learnable filterbanks or recurrent layers for broad hardware compatibility.
- Achieves 0.850 AUC on AVA-Speech with 2.1k parameters and 200ms context—new SOTA for causal VAD on edge.
- Angle-based quantization-aware training (QAT) yields 1–4% improvement over standard QAT, enabling efficient deployment.
Why It Matters
kiloVAD enables always-on voice detection on ultra-low-power chips, extending battery life for smart speakers and wearables.