teensy-vad-v8

teensy-vad-v8 — the 500k scaling ablation

teensy-vad-v8 asks one question: does 10–14× the parameters and 6.6× the training data beat the 42k-parameter v7 champion? No — and the reasons are the most useful result of the campaign. Four ~500k-class models (3-layer GRU-192, 4-layer causal transformer, 2-hidden MLP "a3", each distilled from Silero on 660 hours of prior-balanced mixtures) were trained with quantization-aware fine-tuning and benchmarked on the family's calibrated real-world protocol.

Results (human-labelled real audio, calibrated protocol)

* the numpy int8 matmul runtime is unoptimized; the int8 artifact itself is 4× smaller than float (554 KB vs 2.1 MB).

Findings

1. Scale did not beat fit. No v8 model dominates the 42k v7 champion. The GRU-192 and transformer do edge it on AMI rooms (F1 0.919–0.923), but their clean near-mic ranking collapsed (TEN AUC 0.73–0.78 vs 0.89).
2. The context-MLP scaled best. a3-int8 matches v7's TEN AUC (0.8949 vs 0.8934) with the family's best TEN F1 (0.9133) — at 13× its size.
3. QAT won. int8 fine-tuning improved TEN ranking over float (0.8844 → 0.8949) and held AMI. QAT is validated at 500k scale.
4. Recurrence > attention at every tested scale for clean-audio ranking.
5. Under-training is the honest caveat: the compute budget allowed only ~1.8 passes over 660 h (the deep models likely need 5+). Scale may still win — but only with proportional compute.

Files

License & data

Weights: CC BY 4.0 — commercial use with attribution (© 2026 Pankaj Doharey / Metacritical, TeensyVAD by VoxLogic). Data: LibriSpeech 960 h train subsets (CC BY 4.0), MUSAN (CC BY 4.0), AMI ambience (CC BY 4.0), Silero teacher (MIT). Code: MIT.

Citation

@software{doharey2026teensyvadv8,
  title  = {TeensyVAD-v8: A 500k-Parameter Scaling Ablation for
            Tiny Voice Activity Detection},
  author = {Doharey, Pankaj},
  year   = {2026},
  url    = {https://huggingface.co/Teensy/teensy-vad-v8}
}