🤖 AI Summary
This work addresses the demand for high-accuracy, low-latency voice activity detection (VAD) on resource-constrained edge devices by proposing kiloVAD, a lightweight causal model that relies solely on standard Mel-spectrogram features and a pure convolutional neural network (CNN) architecture. By integrating layer-wise structured pruning, self-distillation, and angle-based quantization-aware training (AQAT), the method significantly enhances post-compression performance without resorting to non-standard or hardware-specific components. The resulting model contains only 2.1k parameters, operates with a 200ms context window, and achieves an AUC of 0.850 on the AVA-Speech benchmark, establishing a new state of the art for deployable causal VAD systems.
📝 Abstract
Voice activity detection (VAD) triggers downstream speech processing in always-on systems under strict memory, latency, and compute constraints. Recent compact models report strong accuracy but rely on components that are not widely supported: learnable filterbanks, recurrent layers, or non-causal post-processing. We propose kiloVAD, designed for embedded inference using standard Mel features, CNN-only layers, and tunable context/spectral parameters. We introduce per-layer structured pruning with self-distillation and angle-based quantization-aware training (QAT) that outperforms standard QAT by 1-4%. Evaluated per-frame under causal conditions, kiloVAD achieves 0.850 AUC on AVA-Speech with 2.1 k parameters and 200 ms context, establishing a new state of the art for causal, deployment-ready VAD.