"Go to sleep, Yossi": microWakeWord stop model (v1)

A dismiss phrase for the Hey Yossi voice assistant, used as Linux Voice Assistant's stop word (--stop-model go_to_sleep_yossi / env STOP_MODEL). It is armed while a reply plays, a timer rings or an announcement runs, and hearing it stops playback.

Files: go_to_sleep_yossi.tflite (streaming, quantized) and go_to_sleep_yossi.json (manifest, cutoff 0.97, window 5).

Training (2026-11-22)

  • Recipe: the lean microWakeWord one (kahrendt/microWakeWord 4665173), mixednet, 10,000 steps, on a Modal T4.
  • Positives: 69 real takes of "Go to sleep, Yossi" from danielrosehill/Hey-Yossi-Stop-Word (revision 8a3dc33), each augmented 50 times. No TTS.
  • Negatives: 40 near-miss takes; 273 "Hey Yossi" takes from danielrosehill/Wake-Words (c0ee33c), so the stop model stays quiet on the wake word; and kahrendt's speech, dinner_party and no_speech sets.
  • Speaker: Daniel only so far. Hannah's takes come next.

Evaluation (held-out test split, cutoff 0.97, window 5)

Set n Result
"Go to sleep, Yossi" 15 recall 0.933 (one annoyed take missed, score 0.52)
"Hey Yossi" (wake word) 55 0 false accepts (max score 0.44)
Near-miss phrases 5 3 fire: "Go to sleep, Josie" (ร—2), "Go to sleep, Yoni". "Time to sleep, Yossi" scores 0.93
Ambient (dinner_party_eval) about 5.3 h 0.19 false accepts/h

Known limit: it hears "Go to sleep, " as the phrase. For a stop word that only listens during playback, that is mostly harmless.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Datasets used to train danielrosehill/Hey-Yossi-Stop-Word-microWakeWord