“zenni. . hold” wake word model

Open-source openWakeWord model for the wake word “zenni. . hold”, trained on openwakeword.com from synthetic speech (not recorded). It runs fully on-device with the openWakeWord runtime: Home Assistant (openWakeWord add-on), Rhasspy, OpenVoiceOS and any Python project.

Model facts

Wake word: “zenni. . hold”
Trained: 2026-03-11 · pipeline V1
Network: 64 × 3 layers, 75,000 training steps
Recall, with background noise and room reverb: 18.2%
False activations on the benchmark set: 29.25 per hour
Files: ONNX (405 KB), TFLite (411 KB)
Downloads: 1

Recall is measured on 500 synthetic test utterances; false activations on about 10 hours of near-miss phrases, noise and read speech, one activation per frame at or above the threshold with a 2-second refractory period.

Download

Sign in on this page to download the model as ONNX (405 KB) or TFLite (411 KB). Downloads are free; models are licensed for personal and non-commercial use by default (commercial license).

Use it in Home Assistant

Download the TFLite file and copy it to the /share/openwakeword folder of your Home Assistant server.
Install the openWakeWord add-on and restart it so it picks up the new file.
In Settings → Voice assistants, open your assistant and choose “zenni. . hold” as the wake word.
Say “zenni. . hold” to your satellite or to the Home Assistant Voice Preview Edition.

Use it in Python

from openwakeword.model import Model

model = Model(wakeword_models=["zenni._._hold_20260311_103755.onnx"], inference_framework="onnx")
# feed 16 kHz mono int16 audio in 1280-sample frames; a score above the threshold is an activation
scores = model.predict(frame)

Similar wake words in the library

“Hey Nova” — 90.4% recall
“Pico” — 44.6% recall
“hey bro” — 81.4% recall
“Hey Jarvis” — 69.8% recall
“Jarvis” — 46.2% recall
“Jarvis” — 36.8% recall

Custom wake words are trained in the Training Center; the library lists every public model with the same measurements.