Mayank022/voice-activity-detection-marblenet

Small streaming voice activity detection model trained from scratch.

Files

• best.pt: best PyTorch checkpoint by validation F1.
• last.pt: final PyTorch checkpoint.
• model.onnx: exported ONNX model for deployment.
• config.resolved.yaml: resolved training config.
• metrics.jsonl: training and validation metrics.
• product_gold_dev_eval.json: full eval output.
• eval_summary.json: eval metrics without per-file predictions.

Training

• Run name: vad_marblenet_4x2x96_a100_10_20h
• Best checkpoint step: 3996
• Best validation F1: 0.9876853606868101

Frame Metrics

• Precision: 1.0
• Recall: 0.9744400540628545
• F1: 0.9870545849773711
• ROC-AUC: 0.0
• PR-AUC: 0.9999958651736498
• EER: 0.0

Product Metrics

• Missed utterance rate: 0.0
• False segment rate: 0.0
• Front-end clipping p95 ms: 50.0
• Endpoint latency p95 ms: 0.0

Caveats

The default validation manifest is based on LibriSpeech dev audio and is useful for early quality checks. For production release, evaluate on hand-labeled audio from the target microphones, rooms, devices, languages, and noise conditions.