MatchBox Keyword Spotting

---

Introduction

This is a small-footprint keyword spotting project. It is using Pytorch. The official MatchBox link is

NeMo/Speech_Commands.ipynb at main · NVIDIA/NeMo

Requirements

1. Python 3.7, 3.8
2. Pytorch 1.8.1 and above

Architecture Diagram

Installation

Install using pip.

pip install -r requirements.txt

Code structure

Below is the file structure of the project.

.
|-- matchbox/
|-- data/
|-- saved_models/
|-- dataloader.py
|-- download_and_extract_noise_file.py
|-- run.sh
|-- testing.py
|-- trainer.py
|-- vad_torch.py
|-- requirements.txt
|-- compute_meand_std.py

• matchbox: matchbox model

• data: consists of json data file. See the json files in the folder to know more details.

  Note: the input dataset is following the json format below.

  {"audio_filepath": "/path/to/audio_wav_0.wav", "duration": time_in_sec_0, 
  "label": target_label_0}

• saved_models: consists of trained model

Training

1. Run the command below to start the training.

   ./run.sh

2. See the results.

Testing

1. Modify the line in run.sh

   python3 testing.py -saved_model save_models/check_point_48_0.9694038245219347

2. See the results.