---
license: apache-2.0
language:
  - en
library_name: openwakeword
tags:
  - wake-word
  - keyword-spotting
  - openwakeword
  - onnx
  - tflite
  - voice-assistant
  - on-device
  - adi
pipeline_tag: audio-classification
---

<img src="https://serve.thelabsource.com/u/VEshKD.png" alt="hey-adi-wakeword" width="800">

# hey-adi-wakeword

Custom-trained **"Hey ADI"** wake word model for the [OpenWakeWord](https://github.com/dscripka/openWakeWord) framework. Built to trigger the **ADI Genesis** offline voice assistant pipeline at theLAB.

Wake phrase: **"hey addie"**

📖 **Build writeup:** [How this wake word was trained](https://thelabsource.com/projects/custom-wake-word.php)

🎥 **Demo video:**

<a href="https://www.youtube.com/watch?v=BSkMgP5AfXw"><img src="https://img.youtube.com/vi/BSkMgP5AfXw/hqdefault.jpg" alt="Hey ADI Wake Word Demo" width="600"></a>

---

## Files

| File | Format | Size | Runtime |
|------|--------|------|---------|
| `hey_addie.onnx` | ONNX | 206 kB | onnxruntime |
| `hey_addie.tflite` | LiteRT (TFLite) | 207 kB | tflite-runtime |

Both files are functionally equivalent — pick the runtime that fits your deployment target. ONNX for desktop/server, TFLite for edge/embedded.

---

## Pipeline

This wake word is the front door to a fully offline voice stack:

OpenWakeWord ("hey addie")  →  Parakeet STT  →  Ollama (advanced-data-intelligence:e4b)  →  OmniVoice TTS (adi_ref voice)

No cloud dependencies — wake detection through response generation runs entirely on local hardware.

---

## Usage

### ONNX (onnxruntime)

```python
from openwakeword.model import Model

model = Model(
    wakeword_models=["hey_addie.onnx"],
    inference_framework="onnx",
)

# frame = 16kHz, 16-bit mono PCM audio chunk (1280 samples / 80ms)
prediction = model.predict(frame)
score = prediction["hey_addie"]

if score > 0.5:
    print("Wake word detected")
```

### TFLite (tflite-runtime)

```python
from openwakeword.model import Model

model = Model(
    wakeword_models=["hey_addie.tflite"],
    inference_framework="tflite",
)

prediction = model.predict(frame)
score = prediction["hey_addie"]
```

### Audio requirements

- **Sample rate:** 16 kHz
- **Channels:** mono
- **Format:** 16-bit PCM
- **Frame size:** 1280 samples (80 ms) per `predict()` call

---

## Threshold tuning

| Threshold | Behavior |
|-----------|----------|
| `0.3` | More sensitive — more false activations |
| `0.5` | Balanced (recommended starting point) |
| `0.7` | Stricter — fewer false activations, may miss quiet/distant speech |

Tune to your microphone and room acoustics.

---

## Intended use

On-device wake word detection for offline voice assistants. Designed for the ADI Genesis stack but works with any OpenWakeWord-compatible application.

## Limitations

- Trained for the English phrase "hey addie" — will not generalize to other phrases.
- Performance varies with microphone quality, background noise, and distance.
- Single-phrase model; pair with additional wake word models if you need multiple triggers.

---

*Built in theLAB — Learning. Algorithms. Breakthroughs.* 🔬