Soulcreek2/speechkit-wakeword-models

SpeechKit Wake-Word Models

Open-source wake-phrase ONNX models trained for SpeechKit via livekit-wakeword (Apache-2.0).

Phrases

File	Phrase	Optimal Threshold	FPPH/h	Recall
hey_quby.onnx	"Hey Quby" (Cubi / Kubi)	0.22	0.06	96.3%
hey_computer.onnx	"Hey Computer"	0.10	0.00	99.7%
hey_jarvis.onnx	"Hey Jarvis"	0.45	0.06	96.2%
hey_mira.onnx	"Hey Mira"	0.08	0.11	99.6%
hey_kombify.onnx	"Hey Kombify"	0.55	0.00	92.4%

Shared upstream

File	Source
melspectrogram.onnx	openWakeWord (Apache-2.0)
embedding_model.onnx	openWakeWord (Apache-2.0)

Pipeline

Standard openWakeWord 3-stage:

1. Mel: [1, 32000] -> [1, 1, 197, 32] (2s audio @ 16 kHz S16 mono)
2. Embedding (sliding 76-frame window, stride 8): -> 16 embeddings of [96]
3. Prediction: [1, 16, 96] -> [1, 1] score

License: Apache-2.0 for the trained phrase heads and for the upstream openWakeWord melspectrogram / embedding models (we use the Apache-2.0 upstream copies, not the CC-BY-NC-SA pre-trained phrase models).

SpeechKit integration

In SpeechKit, set [wakeword] phrase_id = "hey_quby" (or another id) in config.toml and toggle Wake-word on in Settings. The dashboard download manager will fetch the matching files from this repo automatically.

Compatible with any openWakeWord-style runtime (Home Assistant, custom inference loops, etc.) — drop the three files into the runtime's models directory.