teensy-vad-v7 — the tiny recurrent generation

teensy-vad-v7 replaces the family's MLP backbone with a single-layer GRU (96 units) and becomes the best teensy generation on every real-world metric — while also being the fastest model in family history.

Same numpy-only runtime philosophy, same 8 kHz native pipeline, same CC BY 4.0 commercial-safe data (LibriSpeech-100 + MUSAN + AMI ambience), distilled from Silero (MIT). Training runs in torch (scripts/train_rnn.py); the deployed inference path is the verified numpy GRU in teensyvad/rnn.py (max |Δp| vs torch: 1.8e-07 over 250 streamed frames).

Why recurrent

The v6 capacity ablation measured a hard ceiling: widening the 100 ms-window MLP past 80k parameters lowers TEN AUC while raising latency. The missing ingredient was memory — a GRU carries state across the whole call instead of a fixed window, which is precisely what the 1.77M-parameter Silero teacher has and the MLP family could never add by widening.

Results (human-labelled real audio, family protocol)

model	params	KB	TEN F1*	TEN AUC	AMI F1	AMI AUC	µs/20ms
teensy-v7-GRU96	41,809	169	0.8992	0.8934	0.9153	0.9182	37.9
teensy-v6-a2 (best MLP)	49,249	204	0.9081	0.8870	0.8822	0.8726	66
teensy-v5-80k	80,373	321	0.9016	0.8877	0.8845	0.8622	63
Silero VAD (1.77M)	1,774,000	2200	0.9381	0.9519	0.7136	0.8938	94
WebRTC VAD	~6k	~50	n/a	n/a	0.8419	0.7602	2
Energy VAD	—	—	—	0.6702	0.5920	0.6578	7

* TEN at best-F1 threshold; AMI at AMI-dev-calibrated thresholds (thr_hi 0.08), identical protocol for every system.

Highlights

• Beats the entire MLP family on all four real-world metrics — at 42k parameters (half of v6-a2) and 1.7× faster (37.9 µs vs 66 µs).
• Beats Silero on both AMI metrics (F1 0.9153 vs 0.7136; AUC 0.9182 vs 0.8938) — real multi-party rooms, where VAD actually has to work.
• TEN AUC 0.8934 closes further toward Silero's 0.9519; clean near-mic ranking remains Silero's stronghold (recurrent teacher, 36× parameters, orders of magnitude more training data).

Documented negative results (v7 ablations)

Full transparency — two scaling attempts did not beat this checkpoint:

• Capacity: widening the v5 MLP 80k → 200k params lowers TEN AUC (0.8877 → 0.8849) while latency grows 19%. The 100 ms window saturates.
• Naive data scaling: retraining the GRU on 360 h LibriSpeech (4.3× data) with the standard mixture recipe lowers TEN AUC to 0.8586 — the longer utterances push the training prior to 91 % speech frames and the model over-predicts. Prior-balanced mixtures are the follow-up.

Architecture ablation: tiny causal transformer

The same recipe also trained a Kyutai-style tiny causal transformer (d_model 64 × 3 layers, 4 heads, 250-frame = 2.5 s attention window, 120,753 params — teensy-v7-tt.npz in this repo):

model	params	TEN F1*	TEN AUC	AMI F1	AMI AUC	µs/20ms
teensy-v7-GRU96	41,809	0.8992	0.8934	0.9153	0.9182	37.9
teensy-v7-tt (transformer)	120,753	0.8738	0.7403	0.9224	0.9034	34.4

Take-away: the transformer posts the family's best AMI F1 (0.9224) and the lowest latency (34.4 µs), but its windowed-attention ranking degrades sharply on clean near-mic (TEN AUC 0.7403) — the GRU's lossy but unbounded state remains the balanced champion. Attention wins rooms; recurrence wins ranking.

Files

file	use
teensy-v7-gru96.npz	numpy runtime weights + calibrated thresholds (169 KB)
chart_v7.png	comparison chart

Quick start (numpy only):

from teensyvad.rnn import TinyGRU
vad = TinyGRU.load("teensy-v7-gru96.npz")
vad.reset_state(1)
for frame in mel_frames:              # 40-dim log-mel, 10 ms hop @ 8 kHz
    p = vad.step(frame)               # P(speech), state persists

Thresholds ship in the metadata, calibrated on AMI dev meetings (thr_hi 0.08; use thr_hi 0.45 for close-mic telephony).

License & data

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

Limitations

English speech; no music in training; GRU context is unbounded left but causal (no lookahead); no AEC. Silero retains clean near-mic ranking.

Citation

@software{doharey2026teensyvadv7,
  title  = {TeensyVAD-v7: A 42k-Parameter Recurrent Voice Activity
            Detector for Telephony},
  author = {Doharey, Pankaj},
  year   = {2026},
  url    = {https://huggingface.co/Teensy/teensy-vad-v7}
}