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

2026/07/28 by Stephen Bauer, Sheila Seidel, Shanza Iftikhar +2
#eess.AS #cs.LG

paper · pdf

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.

Citations

• SincQDR-VAD: A Noise-Robust Voice Activity Detection Framework Leveraging Learnable Filters and Ranking-Aware Optimization
• Accurate and Structured Pruning for Efficient Automatic Speech Recognition
• Feature Affinity Assisted Knowledge Distillation and Quantization of Deep Neural Networks on Label-Free Data
• DepGraph: Towards Any Structural Pruning
• SG-VAD: Stochastic Gates Based Speech Activity Detection
• MarbleNet: Deep 1D Time-Channel Separable Convolutional Neural Network for Voice Activity Detection
• The INTERSPEECH 2020 Deep Noise Suppression Challenge: Datasets, Subjective Testing Framework, and Challenge Results
• Neural Networks Weights Quantization: Target None-retraining Ternary (TNT)
• pyannote.audio: neural building blocks for speaker diarization
• Optuna: A Next-generation Hyperparameter Optimization Framework
• Speech Model Pre-training for End-to-End Spoken Language Understanding
• AVA-Speech: A Densely Labeled Dataset of Speech Activity in Movies
• CosFace: Large Margin Cosine Loss for Deep Face Recognition
• SphereFace: Deep Hypersphere Embedding for Face Recognition
• Soft-to-Hard Vector Quantization for End-to-End Learning Compressible Representations
• The High-Dimensional Geometry of Binary Neural Networks
• Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations

Related

• Personal VAD: Speaker-Conditioned Voice Activity Detection
• Advancing VAD Systems Based on Multi-Task Learning with Improved Model Structures
• LibriVAD: A Scalable Open Dataset with Deep Learning Benchmarks for Voice Activity Detection
• SG-VAD: Stochastic Gates Based Speech Activity Detection
• Attention Is Not Always the Answer: Optimizing Voice Activity Detection with Simple Feature Fusion