Three Wake-Word Engines You Already Had, as OVOS Plugins
JarbasAl
OVOS Contributor
Three Wake-Word Engines You Already Had, as OVOS Plugins
A wake-word engine is the part of a voice assistant that listens all day for one phrase. Several free engines exist, each with its own model format and runtime, and each is only useful to an OVOS device once something loads it behind the listener's hotwords contract. This post covers three such plugins: ovos-ww-plugin-precise-onnx, ovos-ww-plugin-vosk and ovos-ww-plugin-wakewordlab. They are for people who configure a device. If you run a default image you use the first one without knowing it.
All three register under the opm.wake_word entry-point group of ovos-plugin-manager, so the listener discovers them by the module name in mycroft.conf and nothing else changes.
Precise models on onnxruntime
Precise is the engine the Mycroft community trained its wake words with. Its models were TensorFlow Lite files. ovos-ww-plugin-precise-onnx runs the same models converted to ONNX, with onnxruntime and sonopy for the audio features and no TensorFlow on the device. The precise-lite-models repository carries the converted community set: hey_mycroft, computer, hey_chatterbox, hey_firefox, hey_k9, hey_kit, hey_moxie, hey_scout, jarvis, marvin, sheila, android, christopher, athena and hey_robin, most in both .onnx and .tflite form.
pip install ovos-ww-plugin-precise-onnx
{
"listener": {"wake_word": "hey_mycroft"},
"hotwords": {
"hey_mycroft": {
"module": "ovos-ww-plugin-precise-onnx",
"model": "https://github.com/OpenVoiceOS/precise-lite-models/raw/master/wakewords/en/hey_mycroft.onnx",
"trigger_level": 3,
"sensitivity": 0.5
}
}
}
Version 0.1.0 (stable on PyPI) in a fresh environment downloads the model from that URL and, fed two seconds of silence, stays quiet:
import numpy as np
from ovos_ww_plugin_precise_onnx import PreciseOnnxHotwordPlugin
p = PreciseOnnxHotwordPlugin(key_phrase="hey_mycroft", config={
"model": "https://github.com/OpenVoiceOS/precise-lite-models/raw/master/wakewords/en/hey_mycroft.onnx",
"trigger_level": 3, "sensitivity": 0.5})
for _ in range(20):
p.update(np.zeros(1600, dtype=np.int16).tobytes())
print(p.found_wake_word())
False
Vosk keyword spotting, in any Vosk language
ovos-ww-plugin-vosk takes a different route: it runs the Vosk speech recognizer over short audio chunks in keyword mode and compares the transcript with one or more sample phrases. There is no wake-word model to train. Any phrase in any language Vosk has a model for works, and the plugin downloads the model for the configured language on first use.
pip install ovos-ww-plugin-vosk
{
"listener": {"wake_word": "hey_computer"},
"hotwords": {
"hey_computer": {"module": "ovos-ww-plugin-vosk", "listen": true}
}
}
Four keys matter when a phrase is hard to catch. samples lists the transcripts to accept, so "hey microsoft" can stand in for "hey mycroft". rule sets how a transcript is compared with the samples. full_vocab switches from keyword mode to the model's whole vocabulary. time_between_checks sets the seconds between inferences, from 0.2 to 3. A second entry point, ovos-ww-plugin-vosk-multi, matches several phrases at once.
Version 0.1.10 (stable on PyPI, vosk 0.3.45) builds a two-word grammar for "hey computer" and, on three seconds of silence, returns False. The model download on first run is the slow part.
wakewordlab's compact neural models
wakewordlab is a third-party library of compact neural wake-word models with a Silero voice-activity pre-filter. The detector only runs when someone is speaking. ovos-ww-plugin-wakewordlab wraps it. model is a wake-word slug or a path to a .wkw or .onnx file, threshold is the confidence cut-off, vad and vad_threshold control the pre-filter, and license_key unlocks the library's commercial models. wakewordlab.list_models() prints what is available. Models download on first use into ~/.cache/wakewordlab/models/.
{
"hotwords": {
"hey jarvis": {
"module": "ovos-ww-plugin-wakewordlab",
"model": "hey_jarvis",
"threshold": 0.5,
"vad": true,
"vad_threshold": 0.5
}
}
}
This plugin has GitHub releases (0.0.1a2) and no package on PyPI, so it cannot be installed with pip install ovos-ww-plugin-wakewordlab as its README says. It was not run for this post for that reason. The library itself, wakewordlab 0.1.1, is on PyPI.
Which one
Precise ONNX is the default on OVOS images and the right choice when a community model for your phrase exists. Vosk fits a phrase or a language nobody trained a model for. The cost is a speech-recognition model in memory and a slower check. wakewordlab fits a device that needs voice-activity detection anyway and wants a small neural detector. It stays experimental until it reaches PyPI. A phrase none of them covers is what WakeForge trains, and ovos-ww-plugin-wakeforge runs.
Limits
None of these plugins was tested against a real voice here. Silence proving negative is the check that was run. Vosk's keyword mode is only as good as the model's vocabulary for the phrase. The Precise community models are English-heavy. Bugs go to each plugin's issue tracker on GitHub: precise-onnx, vosk, wakewordlab.
This work is part of the OpenVoiceOS From Beta to Breakthrough milestone, funded through the NGI0 Commons Fund, a fund established by NLnet with financial support from the European Commission's Next Generation Internet programme, under the aegis of DG Communications Networks, Content and Technology under grant agreement No 101135429. Additional funding is made available by the Swiss State Secretariat for Education, Research and Innovation (SERI).
Help Us Build Voice for Everyone
OpenVoiceOS is more than software, it's a mission. If you believe voice assistants should be open, inclusive, and user-controlled, here's how you can help:
- 💸 Donate: Help us fund development, infrastructure, and legal protection.
- 📣 Contribute Open Data: Share voice samples and transcriptions under open licenses.
- 🌍 Translate: Help make OVOS accessible in every language.
We're not building this for profit. We're building it for people. With your support, we can keep voice tech transparent, private, and community-owned.