Project: Label Buddy 2.0: Automated audio-tagging using transfer learning

Mentors: Pantelis Vikatos, Agisilaos Kounelis, Ioannis Sina

Past Mentor: Markos Gogoulos

Contributor: Ioannis Prokopiou

Past Contributor: Ioannis Sina

An annotation tool helps people (without the need for specific knowledge) to mark a segment of an audio file (waveform), an image or text etc. in order to specify the segment’s properties. Annotation tools are used in machine learning applications such as Natural Language Processing (NLP) and Object Detection in order to train machines to identify objects or text. While there is a variety of annotation tools, most of them lack the multi-user feature (multiple users annotating a single project simultaneously) whose implementation is planned in this project. The audio annotation process is usually tedious and time consuming therefore, these tools (annotation tools which provide the multi-user feature) are necessary in order to reduce the effort needed as well as to enhance the quality of annotations. Since in most tasks related to audio classification, speech recognition, music detection etc., machine and deep learning models are trained and evaluated with the use audio that has previously been annotated by humans, the implementation of such a tool will lead to higher accuracy of annotated files, as they will have been annotated by more than one human, providing a more reliable dataset. In effect, multi-user annotation will reduce the possibility of human error e.g. an occasional mistaken labelling of a segment might be pointed out by another annotator.

Deep learning models can be used for annotation and can kickstart your development effort by enabling faster annotation of datasets for AI algorithms. Deep learning models are sensitive to the data used to train them, this makes it hard to train the deep learning models on a specific dataset and deploy them on a different dataset. As a solution, transfer learning for sound could help adapt pretrained models into various datasets. Deep learning models used for annotation can be tuned and improved by retraining these pretrained models based on new datasets.

Already existing annotation tools:

Label Studio: https://github.com/heartexlabs/label-studio

BAT annotation tool: https://github.com/BlaiMelendezCatalan/BAT

Computer Vision Annotation Tool (CVAT): https://github.com/openvinotoolkit/cvat

This project is an enhancement to the previous work that has been done previously. Its goal is to make annotation simple and easy while also providing a well-defined manager-annotator-reviewer framework. The goal of this project is to use Transfer Learning (TL) approaches to make the annotation process easier for the user by offering label predictions. The golas can be devides in two categories of tasks: Machine Learning and Django.

Machine Learning:

• Conduct research for the appropriate model architecture
• Modify the annotation process by integrating the model
• Test the model by providing evaluation metrics

Django:

• Add lazy loading for the audio files: load segments of the file when needed (i.e., YouTube). This will lead to better performance when the audio file is too big.
• Add Django Testing
• Dockerization
• Add documentation
• Add rar file upload functionality - currently, users can only upload zip files (optional)
• UI improvements (optional)

Clone repository and cd to the folder

git clone https://github.com/eellak/gsoc2022-Label-buddy
cd gsoc2021-audio-annotation-tool

Create virtual enviroment

python3 -m venv env

Activate it for Linux

source env/bin/activate

Activate it for Windows

env\Scripts\activate

Install audiowaveform programm following the given steps: https://github.com/bbc/audiowaveform#installation

Install requirements and cd to label_buddy/

pip install -r requirements.txt
cd label_buddy

Make migrations for the Database and all dependencies

python manage.py makemigrations users projects tasks
python manage.py migrate

After the above process create a super user and run server

python manage.py createsuperuser
python manage.py runserver

Visit http://localhost:8000/admin, navigate to users/[your user] and set can_create_projects to true so you can start creating projects.

Visit https://labelbuddy.io/ and sign with the following credentials:

• Username: demo
• Password: labelbuddy123

in order to create projects, upload files and annotate them.

Research & Models Documentation: https://docs.google.com/document/d/10Sd1WPcPpctzXY3lO9p6u4ubNzLg7SEUDKWJv09snqU/edit?usp=sharing

DockerHub Repository: https://hub.docker.com/r/ioannisprokopiou/yoho-training