Keyword Based ASR

Repo for implementation of keyword-based ASR system

Setup

• create virtual enviroment and install requirements from requirements.txt
• NOTE: NeMo toolkit is not supported on Windows, so WSL or UNIX-based OS is required, see the docs or github
• minimal, necessary data is already in the repo, but to reproduce training process and / or test other keywords you need to download full datasets from this link (if the link doesn't work, please contact me via email: [email protected]) and put them in Data directory (check Data README for the structure)
• modify config file to match your setup (all paths with suffix DATA_DIR should be changed to match your setup)

Repository structure

• Data - directory with data
• Utils - directory containing utility scripts and config files
• Models - directory containing trained models' weights
• notebooks in root directory
  • main
    • demo.ipynb - notebook demonstrating dual-model keyword-based speaker recognition system
    • speaker-recognition.ipynb - notebook with speaker recognition model training and evaluation
    • keyword-recognition.ipynb - notebook with keyword spotting model demonstration and evaluation
  • suplementary
    • get-data.ipynb - check audio files metadata
    • visualize-spectrograms.ipynb - visualize spectrograms of audio files
    • play-sound.ipynb - sanity-check audio files