> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cyberwave.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Wake Word — models & compile

> How selected openWakeWord models are downloaded before edge deploy

Wake Word workflows only download the models the author selected — not every file in the openWakeWord catalog.

## Compile-time provision (cloud)

When you **compile** or **edge-sync** a workflow that contains a Wake Word Engine node, `WorkflowCodeAssembler` calls:

```python theme={null}
openwakeword.utils.download_models(model_names=["alexa", ...])
```

Then verifies load with `Model(wakeword_models=[...])` and records `prediction_keys` from `oww_model.models.keys()`.

| Outcome                    | Meaning                                                                                 |
| -------------------------- | --------------------------------------------------------------------------------------- |
| Compile succeeds + warning | `openWakeWord models downloaded for edge: alexa, ...`                                   |
| Compile fails              | `openwakeword` not installed on Django (`ml-wakeword` extra), or download/network error |
| `GET /compile` field       | `openwakeword_models: ["alexa"]` lists normalized model ids                             |

<Warning>
  The Django compile host must have `openwakeword` and `onnxruntime` in `cyberwave-backend/requirements/base.txt`. Rebuild the Django Docker image after adding them. See [Edge workflow dependencies](/feature-reference/workflows/edge-dependencies).
</Warning>

### Compile-server packages

| Package               | Purpose                                |
| --------------------- | -------------------------------------- |
| `openwakeword>=0.6.0` | Model download + load probe at compile |
| `onnxruntime>=1.16.0` | openWakeWord runtime                   |

## Generated worker (edge)

The compiled `wf_<uuid>.py` module begins with:

```python theme={null}
_CW_WW_WORKFLOW_WAKE_WORDS = ['alexa']
# download_models(model_names=_CW_WW_WORKFLOW_WAKE_WORDS)
_cw_ww_ensure_models_downloaded(_CW_WW_WORKFLOW_WAKE_WORDS)
```

Models are fetched again at worker import (before hooks run) and once more when the engine singleton is created.

## Edge-sync metadata

Each Wake Word node contributes to `model_requirements`:

```json theme={null}
{
  "edge_runtime": "openwakeword",
  "edge_package": "ml-wakeword",
  "edge_model_path": "openwakeword/pretrained",
  "wake_word_models": ["alexa"]
}
```

Edge operators use this to confirm the worker image and cache contain the required weights.

## Edge worker image

`cyberwaveos/edge-ml-worker` Dockerfile runs:

```bash theme={null}
python3 -c "import openwakeword; openwakeword.utils.download_models()"
```

at build time (\~20 MB all six built-in models). Workflow-specific `download_models(model_names=[...])` still runs for clarity and custom images.

## Custom images (air-gapped)

```dockerfile theme={null}
RUN pip install openwakeword onnxruntime
RUN python3 -c "import openwakeword; openwakeword.utils.download_models()"
```

Or download only what you need:

```python theme={null}
import openwakeword
openwakeword.utils.download_models(model_names=["alexa", "hey_jarvis"])
```

## Threshold vs compile

**Detection threshold** is a runtime node parameter (default **0.5**, one value for all selected wake words). It is embedded in compiled `wf_*.py` as `threshold=...` on the engine singleton. Changing threshold only requires **save → compile → sync** — no extra model download.

## After changing wake words or threshold

1. Save the Wake Word node in the editor (wake words and/or **Detection Threshold**).
2. **Compile** the workflow (validates models download).
3. **Sync** to the edge environment.
4. Restart the workflow worker container.
5. Confirm logs: `prediction_keys=['alexa']` (not `alexa_v0.1`) and `threshold=0.50` in init / detection lines.