MarbleNet

MarbleNet (ICASSP 2021, Jia/Majumdar/Ginsburg, arXiv 2010.13886) is NVIDIA's answer to the question 'how small can a good neural VAD be?' It's a deep residual network built from blocks of 1D time-channel separable convolution, batch-norm, ReLU and dropout — 88K parameters in the 3x2x64 configuration versus 738K for the CNN-TD baseline it matches. On the multilingual, natural-sounding AVA-speech movie dataset it reports 0.858 TPR at 0.315 FPR (AUROC 0.858) with median smoothing, the best of the compared models. NVIDIA ships it in the NeMo framework, and a multilingual frame-level checkpoint (nvidia/Frame_VAD_Multilingual_MarbleNet_v2.0) is on Hugging Face, reporting ROC-AUC of 95–97% across VoxConverse, AMI, Earnings21, AISHELL-4 and AVA-SPEECH test sets (provider-reported). In practice it's the VAD under the hood of NeMo-based transcription stacks like WhisperS2T-Reborn, which wraps it in a three-tier smoothing/segment-detection system. If you need a neural VAD that runs on a smartwatch-class device, this is the one to reach for — and the paper's ablation studies are genuinely useful guidance for training your own.

• 0.858 TPR at FPR=0.315, AUROC 0.858 on AVA-speech 'All' with 87.5% median overlap, 88K params vs 738K baseline (author-reported, ICASSP 2021)\n- ROC-AUC 92–98% across VoxConverse-test/dev, AMI-test, Earnings21, AISHELL4, CH109, AVA-SPEECH on the HF v2.0 checkpoint (provider-reported)\n- No independent peer-reviewed replication found