voxrt-wake-word 0.1.4

On-device wake-word detection for Linux aarch64 — Raspberry Pi 3/4/5, Zero 2, Jetson, AWS Graviton

voxrt-wake-word

On-device wake-phrase detection for Linux aarch64. Custom Rust inference runtime, ~48K-parameter depthwise-separable convnet, 16 kHz mono PCM in, threshold-crossing events out. Detects the phrase "Hey Assistant".

Runs on Raspberry Pi 3 / 4 / 5 / Zero 2, NVIDIA Jetson, AWS Graviton, and every other aarch64 Linux SBC on a glibc 2.17+ baseline. Measured RTF 0.053 on a Raspberry Pi Zero 2 W — 5.3 % of one A53 core, sustained.

Companion to voxrt-wake-word-android (JitPack) and voxrt-wake-word-ios (SPM). Same runtime, same model, same Detection schema.

Install

pip install voxrt-wake-word

One abi3 wheel covers Python 3.9 / 3.10 / 3.11 / 3.12 / 3.13.

Get the wake-phrase model:

curl -LO https://github.com/VoxRT/voxrt-wake-word-models/releases/download/v0.1.0/voxrt_wake_word.vxrt

Quick start

from voxrt_wake_word import WakeWordEngine

engine = WakeWordEngine.from_path("voxrt_wake_word.vxrt")
engine.threshold = 0.9
engine.cooldown_frames = 100

for chunk in microphone_iter():           # int16 mono @ 16 kHz, any size
    for d in engine.push_pcm_i16(chunk):
        print(f"wake! t={d.timestamp_sec:.3f}s score={d.score:.4f}")

API

class WakeWordEngine:
    @staticmethod
    def from_path(path: str)  -> WakeWordEngine:  ...
    @staticmethod
    def from_bytes(data: bytes) -> WakeWordEngine: ...

    threshold:       float   # default 0.9, sigmoid-space [0, 1]
    cooldown_frames: int     # default 100 (= 1.0 s at 10 ms hop)

    def push_pcm_i16(self, pcm: list[int])   -> list[Detection]: ...
    def push_pcm_f32(self, pcm: list[float]) -> list[Detection]: ...
    def current_score(self) -> float
    def reset(self) -> None

Detection.frame_index: int, Detection.timestamp_sec: float, Detection.score: float — same shape as WakeWordDetection on Android (Kotlin) and iOS (Swift).

Model quality

Test split: 5,240 positive utterances + 6,416 hard-negatives. ROC AUC 0.9966, at the default threshold = 0.9: precision 0.993, recall 0.982, FPR 0.5 %. Full breakdown at thresholds 0.5 / 0.85 / 0.9 / 0.95.

Docs + examples

Full API docs, live-microphone streaming example (via sounddevice), and the equivalent wrappers for Node.js / Go / C / Rust are in the main repository:

https://github.com/VoxRT/voxrt-wake-word-linux

License

• Python wrapper source: Apache-2.0.
• Compiled binary (voxrt_wake_word.abi3.so) + model weights (voxrt_wake_word.vxrt): proprietary — see LICENSE-BINARY inside the wheel or in the repo.

Commercial licensing for custom wake phrases (your brand name, additional languages, multi-phrase): [e-mail] · voxrt.com.