Recognizing Ornaments in Vocal Indian Art Music with Active Annotation

This repository contains the code base used to implement the experiments described in the paper:
Recognizing Ornaments in Vocal Indian Art Music with Active Annotation by Sumit Kumar, Parampreet Singh and Vipul Arora.

• Link to paper (arXiv): (https://arxiv.org/pdf/2505.04419)
• Link to dataset: https://zenodo.org/records/15296955
• For any queries, please contact: [email protected]

🎥 Demo Video

Data Preparation

• dataprep_baseline.ipynb Input:
  Audio folder path, Labels path
  Output:
  Chromagram folder, Labels folder

• audio_chunking.ipynb
  Input:
  Audio folder path, Labels path
  Output:
  Audio chunks folder, Chunks labels folder

• create_chromagrams.ipynb
  Input: Audio chunks folder, Chunks labels folder
  Output:
  Chromagram folder, Labels folder

Training

• train_baseline.ipynb Input: Chromagram folder path, Label folder path (Non-overlapping) Output: Trained baseline model

• train_TCN.ipynb Input: Chromagram folder path, Label folder path (Overlapping) Output: Trained proposed model

• action_seg_model.ipynb Loss function for ED_TCN model

Inference

The pretrained models can be directly used for inference using:

• inference.ipynb

Active Learning Annotation tool

• interface.sh

Demo audio and visualizations

Effect of don't care labelling: Plot 1

Audio 1

Listen / Download Audio

Illustration of false positives in baseline: Plot 2

Audio 2

Listen / Download Audio

Event boundaries modelling: Plot 3

Audio 3

Listen / Download Audio

Long durationed ornaments: Plot 4

Audio 4

Listen / Download Audio