MatchboxNet

A reimplementation of the MatchboxNet architecture for keyword spotting.

The model can be run in the command line with real-time inference by executing run.py (provided you have a working microphone and the necessary dependencies from the requirements.txt installed).

Requirements

Inference

This project uses sounddevice for audio input, which requires PortAudio to be installed.

In Unix, run the following:

sudo apt update
sudo apt install libportaudio2 libportaudiocpp0 portaudio19-dev

Then install the required dependencies via pip. For embedded systems with less RAM (e.g. Raspberry Pi Zero 2 W) you might have to increase have to increase swap space to install larger packages like PyTorch (1GB should be enough).

pip install torch torchaudio sounddevice numpy tqdm

Depending on your microphone you may have to adjust the ENERGY_THRESHOLD in run.py. If the model only outputs SIL, the threshold is likely too high. Vice versa, if you only get false positives, the threshold is likely too low.

Training

WIP

Future goals

• Update train.py to incorporate command line arguments via argparse.