I built an open source VAD that beats Silero, Pyannote, and WebRTC on noisy audio with 93% accuracy — no GPU required

Intermediate
1

Hey everyone,

I built NOVA-VAD, a lightweight explainable Voice Activity Detector that outperforms every major open source VAD on real-world noisy audio.

Benchmark on 100 held-out files from UrbanSound8K (traffic, sirens, jackhammers, construction):

Model Accuracy Lightweight Explainable
WebRTC VAD 58.0% :white_check_mark: :cross_mark:
Pyannote VAD 62.0% :cross_mark: :cross_mark:
Silero VAD 87.0% :cross_mark: :cross_mark:
NOVA-VAD 93.0% :white_check_mark: :white_check_mark:

What makes it different:

  • Pure scikit-learn, no PyTorch or GPU needed
  • Explains every decision with confidence scores and feature importance in plain English
  • Built-in denoiser pipeline
  • Fully retrainable on custom audio data

No existing open source VAD does all three simultaneously.

Would love feedback from this community on the benchmark methodology and approach.

GitHub: GitHub - monishmal3375/nova-vad: Lightweight, explainable VAD that beats Silero and WebRTC on noisy audio. No GPU required. · GitHub

1 Like