v0.1.0

First ship of the keyword-spotting weights.

Model — Streaming Conformer-Medium (d_model=96 × 4 blocks × 8 heads, 636 K params, fp16, 1.28 MB .vxrt v2).

Keywords (14) — yes / no / cancel / play / pause / next / previous / up / down / back / on / off / voxrt / hey_vox.

Format — .vxrt v2 with AES-256-GCM-encrypted tensor blobs + JSON manifest. Consumed unchanged by every VoxRT KWS runtime.

Consumers

• voxrt-kws-linux v0.1.0 — Python, Node.js, Go, C, C++.
• voxrt-kws-browser v0.1.0 — WebAssembly + SIMD128.
• voxrt-kws-{android,ios} — upcoming.

Firing default — sigmoid threshold 0.9, consecutive-3 hold, per-class cooldown 25 emits (1 s at 25 fps).

Numeric parity — bit-identical output vs PyTorch reference across AVX2 (x86_64), NEON (aarch64), and WASM SIMD128 backends (post |Δ| max ≈ 7 × 10⁻⁴).