ericbdc / wakeword

Train a personalised wakeword model thanks to [OpenWakeWord](https://github.com/dscripka/openWakeWord).

Wakeword

Allows to train a personalised wakeword model thanks to OpenWakeWord.

It modifies a bit the provided great python notebook so that you can use it directly as a automatic_model_training.py file (then you can use it without any IPython kernel).

Troubleshoots

For now it is not available on MacOS (because of piper-phonemize library), then you will have to train it on a Linux platform.

Launch on a remote cloud machine

A set of scripts are available in this repository to rapidly configure a remote machine.

Clone this repository on your local machine

git clone [email protected]:ericbdc/wakeword.git
cd wakeword

Get a remote machine

First you need a running Linux platform available, e.g.: AWS EC2 Ubuntu (all runs in ~30min on a t3.2xlarge).

⚠️ Beware that you will need about 32Gb space disk if your chose only 1 hour clips.

In that case, you will generate a Key Pair, name it wakeword and download the wakeword.pem file to your ~/.ssh/ folder.

Check you can connect through ssh, replace user@ip with yours:

ssh -i ~/.ssh/wakeword.pem [email protected]

Configure remote machine

Copy useful scripts

./send_to_remote.sh 'ec2-xx-xx-xx-xxx.eu-west-3.compute.amazonaws.com'

Do the basic setup on the remote machine

ssh -i ~/.ssh/wakeword.pem [email protected] ./setup_remote.sh

Train your model with your wake sentence

Better launch the script in a Linux screen

screen -S wakeword

ssh -i ~/.ssh/wakeword.pem [email protected] python automatic_model_training.py | at now -m

You can detach your screen session by doing Ctrl+A Ctrl+D. If you want to re-attach your screen do screen -R wakeword. To exit your screen do exit and press Enter. You will lose the shell outputs as screen launchs a subshell.

Regenerating the automatic_model_training.py file

Activate the venv and install dev requirements locally

python -m venv venv
source venv/bin/activate
pip install -r requirements-dev.txt

jupyter nbconvert automatic_model_training.ipynb --to python

If you made any change, you will need to copy it again to your remote machine.

Testing the wakeword model

python detect_from_microphone.py --inference_framework=onnx --model_path='my_custom_model/simbad.onnx'