teensy-vad-v6 — context beats capacity

teensy-vad-v6 is the sixth generation of the TeensyVAD family. Instead of adding parameters (which our measured capacity ablation shows hurts past 80k), v6 widens the model's temporal context from 100 ms to 250 ms and adds hard-example mining — and every v6 variant at ≤49k params outperforms the corresponding v5 checkpoint on the real-world benchmarks.

Same numpy-only runtime, same 8 kHz native pipeline, same CC BY 4.0 commercial-safe training data (LibriSpeech + MUSAN + AMI ambience, Silero teacher) as v5.

Ablation: context vs capacity (TEN VAD set ROC-AUC)

variant	params	context	mining	TEN F1*	TEN AUC	AMI F1	AMI AUC	µs/20ms
v5-20k	20,449	100 ms	—	0.8953	0.8760	0.8836	0.8579	64
v5-80k (v5 peak)	80,373	100 ms	—	0.9016	0.8877	0.8845	0.8622	63
v5-150k	148,941	100 ms	—	0.8993	0.8827	0.8861	0.8622	66
v5-200k	198,431	100 ms	—	0.8975	0.8849	0.8849	0.8599	75
v6-a1	24,441	250 ms	—	0.9081	0.8834	0.8842	0.8683	64
v6-a2	49,249	250 ms	—	0.9081	0.8870	0.8822	0.8726	66
v6-c	20,449	100 ms	✓	0.8957	0.8799	0.8853	0.8594	63
Silero VAD (1.77M)	1,774,000	recurrent	—	0.9381	0.9519	0.7136	0.8938	94
WebRTC VAD	~6k	10 ms	—	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.

Findings

• Capacity ceiling, measured: widening the v5 MLP from 80k → 200k params lowers TEN AUC (0.8877 → 0.8849) while latency grows 19%. The 100 ms window saturates at 80k parameters.
• Context is the lever: at 24k params, 250 ms of context beats the family's entire 100 ms capacity curve on TEN F1 (0.9081) and posts the best AMI AUC of any sub-100k model (0.8683).
• Best accuracy/size trade-off in the family: v6-a2 matches the v5 80k peak (TEN AUC 0.8870 vs 0.8877) with 60% fewer parameters, the best AMI AUC of any generation (0.8726), and identical speed.
• Hard-example mining (oversampling Silero-vs-construction disagreement frames) adds ~+0.004 TEN AUC at unchanged size, and ties the family-best AMI F1 (0.8853) at 20k parameters.
• Silero still leads clean near-mic ranking — at 36× the parameters with a recurrent architecture. In real rooms (AMI) every teensy generation since v1-v2 leads its F1 by a wide margin (0.88+ vs 0.714).

Files

file	use
teensy-v6-a1-k25-24k.*	efficiency pick — best TEN F1 in the family at 24k params
teensy-v6-a2-k25-49k.*	recommended default — v5-80k-class accuracy, 60% smaller
teensy-v6-c-20k.*	mining pick — best AMI F1 at 20k params
*.onnx / *-int8.onnx	ONNX float32 / dynamic int8 (22–50 KB), parity-verified

Quick start (numpy only):

from teensyvad import StreamingVAD
vad = StreamingVAD("teensy-v6-a2-k25-49k.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 in the metadata, calibrated on AMI dev meetings (thr_hi 0.10, distant-room profile; use thr_hi 0.45 for close-mic).

License & data

Weights: CC BY 4.0 — commercial use permitted with attribution (© 2026 Pankaj Doharey / Metacritical, TeensyVAD by VoxLogic). Training data unchanged from v5: LibriSpeech train-clean-100 (CC BY 4.0), MUSAN noise (CC BY 4.0), AMI ambience (CC BY 4.0), Silero teacher (MIT). Code: MIT.

Limitations

English speech; no music in training; context is 250 ms (not unbounded — recurrent teachers still hold the clean near-mic crown); no AEC. Full family ablation and protocol: the TeensyResearch paper page and hf_upload/teensy-vad-3/BENCHMARKS.md.

Citation

@software{doharey2026teensyvadv6,
  title  = {TeensyVAD-v6: Context Beats Capacity in Tiny
            Telephony Voice Activity Detection},
  author = {Doharey, Pankaj},
  year   = {2026},
  url    = {https://huggingface.co/Teensy/teensy-vad-v6}
}