teensy-vad-v5 — the commercial-safe retrain

teensy-vad-v5 is the fifth generation of the TeensyVAD family: tiny (20k–100k parameter) voice activity detectors for 8 kHz telephony audio, now trained with a fully commercial-safe data pipeline — the one thing that kept v4 out of commercial deployments.

Same architecture and feature frontend as v4 (log-mel + deltas + 10-frame context → 3-layer MLP, 25/25 ms window, 10 ms hop, 20 mel bands, 80–3800 Hz). The only change is the training data: the environmental noise pool is MUSAN (CC BY 4.0) instead of ESC-50 (CC BY-NC-SA 3.0), so the weights carry no non-commercial restriction.

Results (human-labelled real audio, same protocol as every prior family)

model	params	KB	TEN F1*	TEN AUC	AMI F1	AMI AUC	µs/20ms
teensy-v5 (20k)	20,449	87	0.8953	0.8760	0.8836	0.8579	64
teensy-v5-40k	39,609	162	0.8963	0.8810	0.8853	0.8620	64
teensy-v5-80k	80,373	321	0.9016	0.8877	0.8845	0.8622	63
teensy-v5-100k	99,593	396	0.9008	0.8865	0.8823	0.8596	63
Silero VAD (1.77M)	1,774,000	2200	0.9381	0.9519	0.7136	0.8938	94
WebRTC VAD	—	—	n/a	n/a	0.8419	0.7602	2
Energy VAD	—	—	—	0.6702	0.5920	0.6578	7

* TEN at best-F1 threshold (upper bound); AMI at AMI-dev-calibrated thresholds, identical protocol for every system.

Headlines

• TEN AUC 0.8877 (v5-80k) — new family record, ahead of v4-80k (0.880) and FlashVAD v0.1 (0.882) at the same parameter count.
• AMI F1 0.8853 (v5-40k) is the best room-audio F1 of any teensy generation, at half the size of v4-80k.
• Speed unchanged: ~63–64 µs per 20 ms telephony chunk (single core).
• v5-80k is the recommended default; v5-40k is the size/accuracy sweet spot.

Files

file	use
teensy-v5.npz family (this repo: -20k/-40k/-80k/-100k)	numpy runtime (float32 weights + calibrated thresholds in metadata)
teensy-v5-*.onnx	ONNX float32 (other runtimes/languages)
teensy-v5-*-int8.onnx	ONNX dynamic int8 (22 KB for the 20k model)

Quick start (numpy only):

from teensyvad import StreamingVAD
vad = StreamingVAD("teensy-v5-80k.npz")
for frame in phone_frames:            # 20 ms PCM16LE @ 8 kHz
    for ev in vad.feed(frame):
        print(ev.t, ev.type)          # speech_start / speech_end

Thresholds ship inside the .npz metadata (thr_hi/thr_lo, calibrated on AMI dev meetings; v5 converged on thr_hi 0.10 — the distant-room profile). For close-mic telephony, start at thr_hi 0.45 as in prior generations.

License & data

Weights: CC BY 4.0 — commercial use, redistribution and derivatives permitted with attribution (© 2026 Pankaj Doharey / Metacritical, TeensyVAD by VoxLogic).

Training data is fully commercial-safe:

• Speech: LibriSpeech train-clean-100 (CC BY 4.0)
• Noise: MUSAN (CC BY 4.0) — replacing the ESC-50 CC BY-NC-SA noise of v3/v4
• Ambience: AMI room audio (CC BY 4.0), calibration meetings only
• Teacher labels: Silero VAD (MIT)
• Code: MIT

This is the same architecture and recipe as v4, with the noise pool swapped; the generation is versioned separately so v4 (NC) and v5 (CC BY) lineages remain unambiguous.

Limitations

English speech; no music in training (music reads as activity); 100 ms context (fricatives in noise remain the hardest frames); no AEC — use an echo canceller upstream. Full details: family papers and INTEGRATION.md.

Citation

@software{doharey2026teensyvadv5,
  title  = {TeensyVAD-v5: Commercial-Safe Voice Activity Detection
            for Telephony in 20k--100k Parameters},
  author = {Doharey, Pankaj},
  year   = {2026},
  url    = {https://huggingface.co/Teensy/teensy-vad-v5}
}