VoxRT wake-word model weights

June 4, 2026 · View on GitHub

Pre-compiled wake-phrase detection model weights in .vxrt format, packaged for the VoxRT on-device inference runtime. Detects the phrase "Hey Assistant".

Pair with one of the consumer libraries:

This repo is a thin distribution layer — the actual files live as GitHub Release attachments per version, not in the git tree. The README below is the index.

What is a .vxrt file?

A self-contained, binary, framework-agnostic model file:

  • Topology + weights encoded in a compact format (no ONNX / PyTorch dependency at consume time).
  • AES-256-GCM encrypted at rest, decrypted on load by the runtime (ADR-0023).
  • Versioned — newer .vxrt files may add encoded features that older runtimes refuse to load; check the version compatibility table below.

You don't need to know the format to use it — feed the bytes to VoxrtWakeWordEngine and it Just Works.

Downloads

v0.1.0

FileSizeSHA-256
voxrt_wake_word.vxrt101K9d40bdc132a2ad8e85bd8a28bb49b77c51a7c62f60567222a037e44418510e8f

Compatible with: VoxRT/voxrt-wake-word-ios@v0.1.0, VoxRT/voxrt-wake-word-android@v0.1.0.

Model quality

Test split: 5,240 positive utterances + 6,416 hard-negative utterances (isolated "Hey", isolated "Assistant", competitor wake-words like "Hey Siri", phonetic neighbours, arbitrary speech, non-speech audio). All speakers disjoint from train + val.

  • ROC AUC: 0.9966
  • Average precision (PR AUC): 0.9899

At the recommended deploy threshold of 0.9:

MetricTest value
Precision0.993
Recall0.982
F10.987
FPR0.5 %

See the voxrt-wake-word-android README for the full threshold sweep + per-category false-positive breakdown.

How to use

curl -L -o voxrt_wake_word.vxrt \
  https://github.com/VoxRT/voxrt-wake-word-models/releases/download/v0.1.0/voxrt_wake_word.vxrt
# verify checksum
echo "9d40bdc132a2ad8e85bd8a28bb49b77c51a7c62f60567222a037e44418510e8f  voxrt_wake_word.vxrt" | shasum -a 256 -c

Then bundle (in assets/ for Android, in Xcode resources for iOS) or download at runtime — see the platform repos for code examples.

Provenance + license

The wake-phrase model is trained entirely in-house by Elephant Enterprises LLC (d/b/a VoxRT) on 100% synthetic data — TTS- generated positive utterances ("Hey Assistant" in voice-cloned variants), decomposed hard-negative categories (isolated keywords, competitor wake-words, phonetic rhymes, onset-overlap words, non- speech audio), augmented with real-world RIR + background-noise sources at training time. No upstream model checkpoints, no copyleft or attribution obligations on the weights themselves.

The .vxrt artefact is distributed under the VoxRT proprietary license — see LICENSE. Redistribution as part of an unmodified voxrt-wake-word-{ios,android} library is permitted without per-installation fees; reverse engineering or extracting the unencrypted weights is not. Contact help@voxrt.com for custom phrase models, multi-phrase detection, or licensing terms beyond the redistribution allowance.

Architecture

  • 8-block depthwise-separable Conv1D encoder with dilations [1, 2, 4, 4, 4, 2, 2, 1], 64 channels, ~48K parameters total
  • 64-bin Slaney-norm mel frontend, 16 kHz mono PCM input
  • Global average pooling over 200-frame (2 s) sliding window, linear head, sigmoid output → wake-phrase confidence
  • fp16 weights with AES-GCM at-rest encryption (~100 KB on disk)

Per-frame inference latency on a Snapdragon 662 (Cortex-A73): ~210 µs, giving RTF ≈ 0.021 — well within an always-on power budget.

Versioning

A model release at vX.Y.Z is built to work with VoxRT runtime vX.Y.Z and (typically) all earlier minor versions sharing the same major. Breaking format changes bump the major. See the iOS and Android release notes for the exact compatibility per release.