mlsanityData to 14 Oct 2026

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

arXiv 2607.25870 ↗First posted 28 Jul 2026eess.AS cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

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.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 17 of its 29 references that are in the map. First posted on arXiv 28 Jul 2026.

Builds on

17 of 29 references.

References · 17

  1. 2019Optuna: A Next-generation Hyperparameter Optimization Frameworkcited by 2,501
  2. 2018CosFace: Large Margin Cosine Loss for Deep Face Recognitioncited by 971
  3. 2017SphereFace: Deep Hypersphere Embedding for Face Recognitioncited by 737
  4. 2016Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activationscited by 426
  5. 2023DepGraph: Towards Any Structural Pruningcited by 203
  6. 2019Speech Model Pre-training for End-to-End Spoken Language Understandingcited by 193
  7. 2019pyannote.audio: neural building blocks for speaker diarizationcited by 181
  8. 2017Soft-to-Hard Vector Quantization for End-to-End Learning Compressible Representationscited by 152
  9. 2020The INTERSPEECH 2020 Deep Noise Suppression Challenge: Datasets, Subjective Testing Framework, and Challenge Resultscited by 133
  10. 2017The High-Dimensional Geometry of Binary Neural Networkscited by 22
  11. 2020MarbleNet: Deep 1D Time-Channel Separable Convolutional Neural Network for Voice Activity Detectioncited by 12
  12. 2023Accurate and Structured Pruning for Efficient Automatic Speech Recognitioncited by 11
5 more
  1. 2018AVA-Speech: A Densely Labeled Dataset of Speech Activity in Moviescited by 10
  2. 2022SG-VAD: Stochastic Gates Based Speech Activity Detectioncited by 5
  3. 2023Feature Affinity Assisted Knowledge Distillation and Quantization of Deep Neural Networks on Label-Free Datacited by 5
  4. 2019Neural Networks Weights Quantization: Target None-retraining Ternary (TNT)cited by 3
  5. 2025SincQDR-VAD: A Noise-Robust Voice Activity Detection Framework Leveraging Learnable Filters and Ranking-Aware Optimizationcited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2024CLIP-VAD: Exploiting Vision-Language Models for Voice Activity Detection
  2. 2026Low-Latency State Space Voice Activity Detection with Robust Onset Time Evaluation ↗
  3. 2026Foreground Voice Activity Detection: Learning Speaker Selectivity from Supervision ↗
  4. 2021Voice activity detection in the wild: A data-driven approach using teacher-student training
  5. 2022Joint Speech Activity and Overlap Detection with Multi-Exit Architecture
  6. 2025Edge-ASR: Towards Low-Bit Quantization of Automatic Speech Recognition Models
  7. 2025Adaptive Knowledge Distillation for Device-Directed Speech Detection
  8. 2022Device-Directed Speech Detection: Regularization via Distillation for Weakly-Supervised Models