| # REX Wake-Word Models |
|
|
| Custom wake-word models for the [REX voice assistant](https://github.com/David-Antolick/rex_voice_assistant). |
|
|
| ## Models |
|
|
| - **hey_rex.onnx** - openWakeWord model for the phrase "hey rex". |
| - **hey_rex.tflite** - same model, TFLite runtime version. |
|
|
| Trained with the openWakeWord automatic training pipeline: |
| - 2000 Piper TTS synthetic positives + 329 real recordings from 3 speakers |
| - Background data: FMA music + ACAV100M precomputed features (~2000 hrs) |
| - 10,000 training steps on an RTX 3070 Ti |
|
|
| ## Usage |
|
|
| REX 0.3.2+ downloads this model automatically when wake-word gating is enabled. |
|
|
| For direct use with openWakeWord: |
|
|
| ```python |
| from openwakeword.model import Model |
| oww = Model(wakeword_models=["hey_rex.onnx"], inference_framework="onnx") |
| score = oww.predict(audio_int16_chunk_1280_samples)["hey_rex"] |
| ``` |
|
|
| Audio format: 16 kHz mono int16, fed in chunks of 1280 samples (80 ms). |
| Default detection threshold: 0.5 (raise to ~0.7 to reduce false fires). |
|
|
| ## License |
|
|
| MIT, matching the REX project. Voice samples used in training were collected |
| with explicit consent from contributors for inclusion in this model and its |
| distribution (including any commercial distribution, though none is intended). |
|
|
