Skip to content

v0.1.0

Latest

Choose a tag to compare

@make1986 make1986 released this 21 Jul 11:51

First ship. Full dual-arch (x86_64 + aarch64) SDK across 5 channels — same libvoxrt_kws.so under all of them.

Ships in this release

  • C tarballs: voxrt-kws-0.1.0-{x86_64,aarch64}-linux-gnu.tar.gz (headers + .so + pkg-config + CMake config + 5 examples).
  • PyPI wheels: voxrt_kws-0.1.0-cp39-abi3-manylinux_2_17_{x86_64,aarch64}.whl (pip install voxrt-kws).
  • npm tarball: voxrt-kws-0.1.0.tgz (contains both .node addons, npm install @voxrt/kws).
  • Go module: go get github.com/VoxRT/voxrt-kws-linux/go@v0.1.0.

Model
Streaming Conformer-Medium (d=96, 4 blocks, 636 K params, 1.28 MB .vxrt), 14 keywords: yes / no / cancel / play / pause / next / previous / up / down / back / on / off / voxrt / hey_vox. Download separately from voxrt-kws-models @ v0.1.0.

Perf — 60 s silence, 512-sample chunks, single thread:

  • WSL x86_64 AVX2 + FMA: RTF 1.03 % (97× realtime).
  • Pi Zero 2 W NEON A53 @ 1 GHz: RTF 32.2 % (3.1× realtime).
  • AVX2 and NEON produce bit-identical output vs PyTorch on the 5-sample parity suite.

Hardening — glibc 2.17 baseline (manylinux2014), full RELRO + non-exec stack, no build-id / .comment / .note.gnu.property, linker version-script restricts exports to the voxrt_kws_* 14-symbol public C ABI. AES-256-GCM weight encryption, XOR-obfuscated master key.

Examples shipped — 5 in C/C++ (alsa-mic-quickstart{,-cpp}, ws-detect-server{,-cpp}, cmake-consumer), 3 each in Python / Node / Go (quickstart, alsa-mic-quickstart, ws-detect-server).

Validated end-to-end via public registries on WSL (WebSocket + browser mic) and Pi Zero 2 W (ALSA mic) across all 5 languages.