wakewords 0.3.7

Tools for building wake-word and speech-command datasets and models.

wakewords

Build custom wakeword and command-word datasets from TTS-generated words plus Google Speech Commands.

Quick Start

Create A Project

Initialize the project layout:

uv run wakewords init

This creates data/, background_audio/, config.json, and a project .gitignore entry for downloaded Google Speech Commands data.

Edit config.json and put your wake words in custom_words.

Set Up TTS

The default TTS provider is Cartesia. Set your API key before generating audio:

export CARTESIA_API_KEY=your-api-key

Custom TTS providers can be registered from config.json. See docs/custom-providers.md.

Generate English Data

Generate clean samples for the custom_words in the project config.json using every available English voice:

uv run wakewords generate --lang en --all-voices

Generated audio and metadata are written to the project's data/custom_words.parquet.

Augment The Dataset

Create noisy tempo variants for the generated clean samples:

uv run wakewords augment

By default, augmentation targets about 4000 total samples per word.

Check Data

Print duration and no-speech stats for generated and augmented rows:

uv run wakewords checkdata

Use --generated or --augmented to check only one source type. No-speech sample IDs are written to no-speech.txt in the project root.

Train

Download Google Speech Commands, build manifests, and preview the training run:

uv run wakewords download
uv run wakewords manifest
uv run wakewords train --dry-run

Run training on Linux with NeMo installed:

uv run wakewords train

Training uses NeMo's from_pretrained() by default. To train from a local .nemo file instead, pass --base-model-path.

Export

Export the latest completed training run into a project-level model bundle:

uv run wakewords export --format onnx

This writes models/model.onnx for inference, plus models/last_checkpoint/last.ckpt, models/last_checkpoint/train_config.json, models/labels.json, and models/export_config.json when those source files are available. The checkpoint directory is kept ready for continued training with the original training settings.

Resume from an exported checkpoint bundle with:

uv run wakewords train --from-checkpoint models/last_checkpoint/last.ckpt

That imports the checkpoint into a new runs/<run-name>/ directory before training continues.

Find Outputs

Training artifacts are written under runs/<run-name>/:

• train_config.json
• checkpoints/
• logs/
• models/

The final exported model is written under the models/ directory of that specific training run.

More Details

See docs/USAGE.md for command options, split ratios, augmentation details, cleaning commands, and training notes.

License

Copyright © 2026 Akash Manohar John, under MIT License (See LICENSE file).

Background sounds: The background audio embedded in this pypi package comes from the Google Speech Commands dataset and ships with this library for convenience. This is licensed under the same license as the dataset. The details are in the README.md file inside of the wakewords/google_scd_background_noise dir.