Listening Lab - Annotator

The listening lab is an open-source platform for audio annotation of less vocal and typically more challenging species.

This tool allows you to:

• Upload raw field recordings
• Automatic segment sparse features of interest to reduce processing time
• Train state-of-the-art transformer-based model for your application
• Export annotations and models for use in the field
• Outlier detection

This tool was developed by the University of Canterbury's Listening Lab Bioacoustic Research group https://github.com/listening-lab

Setup

Start frontend using cd frontend then npm start after installing dependences

Start backend server using cd backend then uvicorn main:app --reload

Python dependences in ./backend/environment.yml

Docker

To start a clean build on your local machine run docker-compose up -d after cloning the repository.

Docker Hub

Install the latest images using docker hub docker image pull -a benmcewen/listening-lab

Start the frontend using docker run -p 3000:3000 benmcewen/listening-lab:frontend and the backend using docker run -p 8000:8000 benmcewen/listening-lab:backend

Classification Pipeline

We current use a transformer-based classification model - Audio Spectrogram Transformer (AST). The implementation can be found in ./backend/preprocessing/classifier.py.

Citation

If you find this tool useful, please cite it (journal publication coming soon!)

@article{McEwen2023,
  title = "Listening Lab - A Human-in-the-Loop Approach for Annotation of Sparse Audio Events",
  author = "{McEwen}, Ben and {Soltero}, Kaspar and {Gutschmidt}, Stefanie and {Bainbridge-Smith}, Andrew and {Atlas}, James and {Green}, Richard",
  journal = "unpublished",
  year = 2023
}

Contribute

Feel free to contribute to this project or adapt it for your application.

TODO

• Set the model to automatically update prototypes
• Scale points in point map view
• Testing

Bug fixes

• Prototype removed when label removed and unknown prototype not included at inference