Train custom wake word models with openWakeWord. A granular 13-step pipeline with compatibility patches for torchaudio 2.10+, Piper TTS, and speechbrain. Generates tiny ONNX models (~200 KB) for real-time keyword detection — like building your own "Hey Siri" trigger.

This toolkit automates the entire openWakeWord training process:

1. Synthesizes thousands of speech clips using Piper TTS with varied voices and accents
2. Augments clips with real-world noise, music, and room impulse responses
3. Trains a small DNN classifier optimized for always-on, low-latency detection
4. Exports a tiny ONNX model you can deploy anywhere

The result is a ~200 KB model that runs on CPU in real-time with negligible resource usage.

wsl
nvidia-smi

You should see your GPU listed. If not, update your NVIDIA Windows driver to the latest version.

# From PowerShell (Windows) — cd to the repo first:
cd path\to\openwakeword-trainer
wsl -- bash train.sh

# Or from within WSL2/Linux:
cd /mnt/c/path/to/openwakeword-trainer
bash train.sh

This creates an isolated virtualenv, installs dependencies, downloads datasets, trains the model, and exports the result.

# Enter WSL2 and navigate to the repo
wsl
cd /mnt/c/path/to/openwakeword-trainer

# Create & activate a training venv (use native filesystem, not /mnt/c/)
python3 -m venv ~/.oww-trainer-venv
source ~/.oww-trainer-venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Run the full pipeline
python train_wakeword.py

# Or resume from where you left off
python train_wakeword.py --from augment

1. Copy the example config:

   cp configs/hey_echo.yaml configs/my_word.yaml

2. Edit configs/my_word.yaml:

   model_name: "my_word"
   target_phrase:
     - "hey computer"
   custom_negative_phrases:
     - "hey commuter"
     - "computer"
     - "hey"

3. Train:

   python train_wakeword.py --config configs/my_word.yaml

4. Find your model in export/my_word.onnx (and export/my_word.onnx.data).

The pipeline runs 13 granular steps, each with built-in verification. If any step fails, it stops immediately and tells you exactly how to resume.

If any step fails:

Pipeline stopped.  Fix the issue above, then resume:
  python train_wakeword.py --from <failed-step>

# Full pipeline (all 13 steps)
python train_wakeword.py

# Use a custom config
python train_wakeword.py --config configs/my_word.yaml

# Resume from a specific step
python train_wakeword.py --from augment

# Run exactly one step
python train_wakeword.py --step verify-clips

# Check current state without side effects
python train_wakeword.py --verify-only

# Show all available steps
python train_wakeword.py --list-steps

The export step produces two files that must be kept together:

• hey_echo.onnx — the model graph (~14 KB)
• hey_echo.onnx.data — external weights (~200 KB)

Copy both files to your project. The trained model works with any openWakeWord-compatible runtime:

from openwakeword.model import Model

oww = Model(wakeword_models=["export/hey_echo.onnx"])

# Feed 16 kHz audio frames
prediction = oww.predict(audio_frame)

Or with ONNX Runtime directly:

import onnxruntime as ort
import numpy as np

sess = ort.InferenceSession("export/hey_echo.onnx")
# Input shape: [1, 16, 96] (mel spectrogram features)
result = sess.run(None, {"x": features})

See configs/hey_echo.yaml for a fully commented example. Key settings:

After training, tune the detection threshold for your use case:

This toolkit includes automatic patches for known breaking changes in modern dependency versions:

Patches are applied and verified automatically during the apply-patches step.

After training, reclaim disk space:

rm -rf data/          # ~12 GB of downloaded datasets
rm -rf output/        # intermediate training artifacts

Keep only the export/ directory with your trained model.

This package only has Linux wheels. Make sure you're running inside WSL2, not native Windows.

Update your NVIDIA Windows driver to the latest version. WSL2 CUDA passthrough is included automatically.

Verify CUDA is available: python -c "import torch; print(torch.cuda.is_available())". If False, everything falls back to CPU.

Reduce tts_batch_size in your config (e.g., 25 → 10).

Re-run the script — all downloads are idempotent and resume where they left off.

MIT — see LICENSE.

• openWakeWord by David Scripka
• Piper by Rhasspy for synthetic TTS
• Built with PyTorch, ONNX Runtime, and speechbrain