teensy-vad-v4

teensy-vad-v4 — the 100-hour family

The fourth generation of the teensyvad family: the same readable 3-layer MLP, now trained on the full ~100 h of LibriSpeech train-clean-100 (all 28,539 utterances — 3.5× the v3 dataset, 37.9M teacher-labelled frames), with the complete 20k/40k/80k/100k capacity sweep.

This card is the home of the family; every named size variant lives in this repository.

Quick Start

from teensyvad import OfflineVAD

# Standalone VAD — auto-downloads from the Hub (pip install huggingface_hub)
model = OfflineVAD("Teensy/teensy-vad-v4")
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-v4", model_file="teensy-v4-80k.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-v4")
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.

Why 100 hours

The v3 experiment showed capacity only pays when data scales with it: on 1M frames, 20k→100k params was flat; on 10.7M frames the sweet spot moved to ~80k. v4 completes the curve at 37.9M frames — and this time capacity keeps paying to the largest size tested:

Real-world results (TEN VAD public set + AMI SDM, human labels) are in the comparison section below; the shared protocol lives in BENCHMARKS.md.

Models in this repository (named variants)

Training

• Speech: all 28,539 utterances of LibriSpeech train-clean-100 (~100 h of mixtures; 37,936,872 frames — every window verified from the memmap at load: train 37,936,863 windows)
• Noise: ESC-50 + synthetic 7-talker babble + real AMI room ambience (calibration meetings only), SNR −5 … 20 dB, G.711 µ-law augmentation
• Labels: Silero teacher (hard), memmap/float16 training plumbing (the design matrix would be 60 GB if materialised)
• Architecture identical to teensy-vad-1 — see its card for the exact reimplementable feature spec.

Comparison vs Silero / WebRTC / Energy — and prior families

Shared protocol (human-labelled audio, AMI-dev-calibrated operating points, AUC on raw probabilities — full appendix: BENCHMARKS.md):

* tuned on that set — like-for-like with FlashVAD's published F1 0.889 / AUC 0.882: v4-80k essentially matches FlashVAD's AUC (0.880 vs 0.882) at 2.3× fewer parameters, and leads every prior teensy family on TEN AUC. The int8 QAT variant holds TEN AUC 0.876 in a 28 KB file — within 0.005 of the 80k champion at 1/11th the size (val F1 0.9145, matching float v4).

License & data

Code: MIT. Weights: CC BY-NC-SA 4.0 (ESC-50 CC BY-NC-SA 3.0 noise in training; LibriSpeech CC BY 4.0; AMI CC BY 4.0; Silero teacher MIT). Not for commercial deployment without replacing non-commercial training data.