teensy-vad-2 — distilled from Silero VAD

Same tiny architecture as teensy-vad-1 (20,449 params, 87 KB, pure numpy), but trained on Silero VAD teacher labels instead of construction labels — knowledge distillation from a ~2M-parameter teacher into a 20k-parameter student.

Why: construction labels know where the utterance was placed, not where speech actually starts and stops. Silero relabelled every frame of the training mixtures (soft probabilities, interpolated to the 10 ms grid); the teacher disagreed with construction labels on 11.6 % of frames — 9.2 % were pauses/breathy tails mislabelled "speech". The student learned where speech actually is:

boundary quality (vs Silero teacher, 100 streamed clips)	v1	v2
onset Δ	−432 ms	−172 ms
offset Δ	+718 ms	+478 ms

Quick Start

from teensyvad import OfflineVAD

# Standalone VAD — auto-downloads from the Hub (pip install huggingface_hub)
model = OfflineVAD("Teensy/teensy-vad-2")
result = model.segments("long_audio.wav")
# Returns speech segments: [[start_ms, end_ms], [start_ms, end_ms], ...]
print(result)                    # e.g. [[90, 5150]] — the fsmn-vad convention

Variant selection: OfflineVAD("Teensy/teensy-vad-2", model_file="teensy-v2-80k.npz") (see the variant table below). Local use without the hub: pass a .npz path. teensyvad is pure numpy (single package dir in the project repo) — the only dependency.

Use as Part of an ASR Pipeline

from teensyvad import OfflineVAD
from teensyvad.audio import write_wav

vad = OfflineVAD("Teensy/teensy-vad-2")
segments = vad.segments("meeting_2hours.wav")     # [[start_ms, end_ms], ...]

for i, (start_ms, end_ms) in enumerate(segments):
    seg = vad.slice("meeting_2hours.wav", start_ms, end_ms)  # float32 @ 8 kHz
    write_wav(f"speech_{i}.wav", seg, 8000)
    # text = my_asr.transcribe(f"speech_{i}.wav")  # <- your ASR here

Streaming / telephony (Asterisk AudioSocket): from teensyvad import StreamingVAD — feed 20 ms PCM16LE frames, get speech_start / speech_end events, vad.speech_seconds for talk time; a working server ships as scripts/audiosocket_server.py in the project repo.

Architecture

Identical to v1 (see its card for the exact, reimplementable feature spec): 400→48→24→1 MLP over 10 frames of 20 log-mel+Δ at 8 kHz.

What's in this repo

file	what
teensy-v2.npz	float32 model (hard teacher labels) — the default v2
teensy-v2-qat.npz	QAT int8 version (28 KB, decision agreement 96 % with float)
teensy-v2.onnx	float32 ONNX export
teensy-v2-int8.onnx	dynamic int8 ONNX (22 KB, 0.067 µs/frame batched)

Training

• Same mixtures as v1 (LibriSpeech dev-clean + ESC-50, 0–20 dB SNR, µ-law augmentation), labels from Silero VAD run at 8 kHz
• Two students trained: hard 0/1 labels (this model) and soft probabilities (slightly higher AUC 0.929 vs 0.922 on val; the hard variant produced tighter event boundaries and is the default)
• Event thresholds calibrated at event level on validation

Measured quality

benchmark	frame F1	AUC
synthetic test (vs teacher labels)	0.905	0.923
TEN VAD public set (stored thr)	0.869	0.868
TEN VAD public set (best thr, upper bound)	0.890	—
AMI SDM meetings (AMI-calibrated thr)	0.880	0.848

Teacher (Silero) on the same real sets for scale: TEN F1 0.937 / AUC 0.863; AMI F1 0.714 / AUC 0.772 — the student already out-ranks its teacher on AMI (real rooms) while the teacher stays better calibrated close-mic.

Quantization

PTQ is essentially free on this model (ΔAUC 0.0000, 99.7 % decision agreement); QAT adds +0.9 F1 points over PTQ at identical size. In pure numpy int8 is a size play (~3× smaller) not a speed play — for speed use the ONNX int8 file (real int8 kernels, 0.067 µs/frame batched).

Limitations (honest)

• A student inherits its teacher's biases — Silero's conservatism on sung/tone-like audio transfers.
• English read speech + environmental noise; no babble or room tone (see teensy-vad-3).
• Operating thresholds are domain-specific: v2's real-set numbers above used AMI-calibrated thresholds (0.10); the stored default suits synthetic-event use.

Models in this repository (named variants)

All variants share the architecture and distillation recipe of this card — float variants differ only in hidden-layer size (88/48, 164/88, 200/96); the QAT variant is the 20k net fine-tuned under int8 simulation:

name	file	params	KB	role
teensy-v2	teensy-v2.npz	20,449	87	the default — distilled 20k, tightest boundaries
teensy-v2-40k	teensy-v2-40k.npz	39,609	162	capacity probe
teensy-v2-80k	teensy-v2-80k.npz	80,373	321	capacity probe
teensy-v2-100k	teensy-v2-100k.npz	99,593	396	capacity probe
teensy-v2-qat	teensy-v2-qat.npz	20,449	28	int8 QAT — smallest accurate artifact

ONNX exports (float32 + dynamic int8) are provided for the 20k model.

Accuracy & speed of every variant

(Measured on the shared protocol — TEN VAD public set + AMI SDM meetings, human labels, AMI-dev-calibrated operating points, AUC on raw probabilities; see BENCHMARKS.md. Accuracy is flat from 20k→100k on the 1M-frame training set — the distilled 20k is the family's sweet spot; scale data instead, see teensy-vad-3.)

Comparison vs Energy / WebRTC / Silero

Same protocol for every system, human-labelled real audio, operating points calibrated on held-out AMI dev meetings (full appendix with charts and per-size results: teensy-vad-3/BENCHMARKS.md):

	teensy-v2	Silero VAD	WebRTC VAD	Energy VAD
params	20,449	1,774,000	~6k (C)	—
TEN VAD set — F1 (best thr*)	0.890	0.938	n/a	—
TEN VAD set — AUC	0.868	0.952	n/a	0.670
AMI SDM — F1 (calibrated)	0.880	0.714	0.842	0.592
AMI SDM — AUC	0.848	0.894	0.760	0.658
µs / 20 ms chunk	63	89	2	7

* like-for-like with FlashVAD's published TEN numbers (also threshold-tuned on that set): FlashVAD F1 0.889 / AUC 0.882 — teensy-v2 matches the F1 with 2.3× fewer parameters.

Capacity note: this family was also trained at 40k/80k/100k params (teensy-v2-{40k,80k,100k}.npz) — like v1, accuracy is flat across sizes on the 1M-frame training set. The distilled 20k model is the family's sweet spot; scale the data instead (see teensy-vad-3).

License & data

Code: MIT. Weights: CC BY-NC-SA 4.0 (ESC-50 training noise is CC BY-NC-SA 3.0; speech LibriSpeech CC BY 4.0; teacher Silero VAD is MIT). Not for commercial deployment without retraining on permissive noise data.