wakeforge

wakeforge is a research framework for wake-word detection. It implements architectures, losses, and featurizers from the published wake-word literature, so you can reproduce a paper's result or compare methods side by side. Some included architectures exist to reproduce a published result, not because they are the best default — the per-architecture pages under docs/reference/ say which is which. Train, evaluate, and export lightweight on-device detectors that run anywhere from an ESP32 to a GPU server. Every component exports to ONNX; production inference requires only onnxruntime and numpy — no PyTorch at runtime.

An easy training run is not the same thing as a good detector. Loss going down and a quickstart command exiting 0 prove the pipeline works, not that the result is safe to ship — see docs/guides/expectations.md for what a deployable wake word actually looks like in false-accept and false-reject numbers, and how much data each tier needs to get there.

What is a wake word?

A short phrase ("hey jarvis", "computer", "alexa") that a device listens for continuously. When detected, downstream STT/NLU runs. A useful detector must run on tiny hardware (sub-100 KB, <10 % CPU, no internet), tolerate noise and distance, almost never false-fire, and trigger reliably when spoken. wakeforge is the toolchain that builds such a detector from a single phrase — synthesise data, train, evaluate, export, deploy. What counts as "almost never" and "reliably" in numbers is in docs/guides/expectations.md, not repeated here.

Who is this for?

Highlights

• Single-string-to-ONNX quickstart — train_from_wakeword("hey jarvis", out) produces a deployable model.
• 11 built-in featurizers (+ enrichment wrappers) × 15 classifier heads × 15 losses — a real research surface.
• Genetic + Bayesian HP search with island-model parallelism, adaptive mutation, two-stage refinement.
• Synthetic datagen — TTS + pure-ONNX voice conversion (voiceclonnx) to bootstrap a dataset from zero recordings.
• Hard-negative mining and infinite training for industrial-scale negative pools.
• ONNX-first: featurizer and head export cleanly; no CUDA-only kernels.
• Hardware tiers from esp32_nano (sub-1 KB int8) to hubert_medium.

Honest trade-offs

• CPU training works for small tiers; a mid-range GPU is the best UX for larger ones.
• Synthetic data is great for smoke-testing — production still needs real far-field recordings.
• ONNX-export is mandatory; non-traceable components (custom CUDA kernels, dynamic control flow) are out of scope. See docs/internals/known_issues.md.
• SSL featurizers (HuBERT, Wav2Vec2-BERT) are used as pre-exported ONNX and held frozen during downstream training — guarantees train/inference parity but limits adaptation.

Install

# Published package (import name stays ww_trainer):
pip install --pre "wakeforge[datagen,torchcodec]"

For development, clone the repo and install it editable instead:

# Core library + tests
uv pip install -e ".[dev]"

# Quickstart / datagen needs TTS plugins + HF datasets + an audio codec:
uv pip install -e ".[dev,datagen,torchcodec]"

Optional extras (sweep, transformers, mlflow, datagen, vc, mic, viz, markov, ocsvm, torchcodec) — see docs/faq.md.

A default quickstart run needs ≈ 6–8 GB disk and ~5 GB download (or ~1.5 GB with --no-augmentation-data). Voice cloning via --vc-refs adds the per-engine ONNX weights downloaded on first use from the HuggingFace Hub (size varies by voiceclonnx engine). Full per-dataset budget: docs/getting_started/requirements.md.

60-second quickstart

ww_trainer-quickstart --wake-word "hey jarvis" --output-dir ./hey_jarvis

Or in Python:

from ww_trainer.quickstart import train_from_wakeword
result = train_from_wakeword("hey jarvis", "./hey_jarvis",
                             tier="small", epochs=50)
print(result.best_onnx_path, result.metrics)

Output: best_f1_featurizer.onnx + best_f1.onnx under ./hey_jarvis/model/. Load both with OnnxWakeWordInferencer — docs/guides/inference.md.

Documentation

Everything lives in docs/. Start with:

• docs/learning_path.md — zero-to-hero curriculum with literature anchors
• docs/faq.md — topic-ordered Q&A in 15 sections
• docs/index.md — full documentation index
• examples/README.md — 43 runnable examples

Contributing

Issues and pull requests welcome on the dev branch. Tests live in test/; run with uv run pytest.

Citation

@software{wakeforge,
  title  = {wakeforge: a research framework for on-device wake-word detection},
  author = {TigreGotico contributors},
  year   = {2026},
  url    = {https://github.com/TigreGotico/wakeforge},
  note   = {Funded by NGI0 Commons Fund / NLnet, grant 101135429}
}

Credits

Developed by TigreGótico for OpenVoiceOS.

This project was funded through the NGI0 Commons Fund, a fund established by NLnet with financial support from the European Commission's Next Generation Internet programme, under the aegis of DG Communications Networks, Content and Technology under grant agreement No 101135429.

License

Apache 2.0