Wakeword Data Collector is a prototype CLI to record wakewords, non-wakewords, and background noise written in Python. Think of it as your wakeword data collection recipe for creating bullet proof wakeword models. It's part of the Secret Sauce AI Wakeword Project.
The Wakeword Data Collector records wave files with a sample rate of 16000 for two main categories of data:
- Wakewords (ie 'hey Jarvis')
audio/wake-word/- Wakeword variations (ie saying 'hey jarvis' further/closer to the mic, faster, or slower)
audio/wake-word/variations/
- Wakeword variations (ie saying 'hey jarvis' further/closer to the mic, faster, or slower)
- Not-wakewords
audio/not-wake-word/- background noise recordings
audio/not-wake-word/background/ - syllables (ie 'hey, jar, vis')
audio/not-wake-word/parts/ - combinations of syllable permutations (ie 'hey jar', 'jarvis')
audio/not-wake-word/parts/ - other longer recordings: recording the TV and a natural conversation
audio/not-wake-word/
- background noise recordings
You may have to install the pyaudio dependancy, ie:
sudo apt-get install portaudio19-dev.
As usual once you have cloned the repo, it is recommended to create and activate python virtual environment to install the requirements.
python3 -m venv .venv
source .venv/bin/activate
pip3 install -e .Simply run wakeword_collect in your console and follow the instructions.
wakeword_collect
For a first time user, it is highly recommended to do a full data collection of all steps (besides 3. Non-wake-word recordings which is optional) to ensure a production quality wakeword.
1. First set of 16 wakeword recordings2. Wakeword syllable and syllable permutation recordings4. Second set of 16 wakeword recordings5. Wakeword variations6. Third set of 16 wakeword recordings7. First random conversation recordings8. Second random TV recording9. Second rantom TV recording- The background recordings taken throughout the sessions will be in
audio/not-wake-word/background/
If you are doing a data collection to add another user it is recommended to record:
1. First set of 16 wakeword recordings7. First random conversation recording- All of the background recordings from this session in
audio/not-wake-word/background/ 2. Wakeword syllable and syllable permutation recordingsis optional, but it can help
- It is recommended to use your production audio hardware and location to collect the samples.
- It is best to use a wakeword with at least three syllables.
- Make sure to check each recording, it is always possible something went wrong. Even one bad or missing recording can be the difference between a bullet proof and a craptastic wakeword model.
- After collecting your data, it is highly recommended to run the Precise Wakeword Model Maker.
Although Secret Sauce AI is always about collaboration and community, special thanks should go to Dan "The Man" Borufka for his support since day one. Thanks, Dan!