⠀⠀⠀⣀⣀⣀⣤⠶⠚⠛⠛⠓⠶⣤⣀⣀⣀⠀⠀⠀
⠀⠈⠉⠉⠉⠁⢀⡤⠞⠋⠙⠳⢤⡀⠈⠉⠉⠉⠁⠀
⢠⣤⡤⠴⠶⠚⠉⠀⣀⣤⣤⣀⠀⠉⠓⠶⠦⢤⣤⡄ S E N N E T.
⠀⠀⠀⠀⢀⣠⡴⠛⠁⠀⠀⠈⠛⢦⣄⡀⠀⠀⠀⠀ Talk to your coding agents.
⠀⠛⠛⠋⠉⠁⠀⠀⣀⣀⣀⣀⠀⠀⠈⠉⠙⠛⠛⠀
⠀⠀⢤⣤⣤⠴⠖⠛⠉⠁⠈⠉⠛⠲⠦⣤⣤⡤⠀⠀
"Hey Sennet" wake word
An openWakeWord model for the phrase "hey sennet"
(said like "senate"), the wake word of Sennet. It's a
small ONNX classifier (~200 KB) that runs on openWakeWord's shared mel spectrogram and speech
embedding models. Sennet runs it in the browser; sennet setup downloads it if you choose it.
Files
hey_sennet_v2.onnx: the current model. Older versions stay here so older installs still download theirs.
Training
Trained with openWakeWord's training code (Sennet's training/wakeword):
- 100,000 spoken "hey sennet" clips from piper-sample-generator's LibriTTS-R voice (904 speakers mixed), and as many sound-alike phrases as negatives ("hey sonnet", "hey senator", "senate"…), a quarter of them the close ones.
- Room echo, background noise and music mixed in (AudioSet, FMA, MIT room impulse responses).
- openWakeWord's precomputed ACAV100M features (~2,000 hours) as further negatives, tuned to aim for 0.2 false wakes an hour.
On held-out synthetic clips with noise and echo, at a 0.5 threshold, v2 has a recall of 0.28, 0.0 false wakes an hour and an accuracy of 0.64. That reads lower than it feels with a real voice: in use it wakes well to "hey sennet", and not to "hey senator" or "okay senator".
Licence
CC BY-NC-SA 4.0, for non-commercial use only, as openWakeWord's own pre-trained phrases are: the training data includes datasets with non-commercial or uncertain licences.