Name That Tune: Jaka to kolęda?

Overview

The repository contains a project focused on classifying Polish Christmas Carols (kolędy) based on chroma features extracted from melodies. The project structure includes the following:

• Carol_Classifier: Main directory containing the Python scripts and Jupyter notebooks related to the project.
  • Carol_Classification-Melody_Extracts.ipynb: Jupyter notebook for the classification of songs with highlighted main melody.
  • Carol_Classification-Original_Songs.ipynb: Jupyter notebook for the classification of original songs.
  • ChromaCoverId: Component related to chroma feature extraction cloned from here.
  • Data_prep.ipynb: Jupyter notebook focusing on data preparation.
  • Melody_extractor.ipynb: Jupyter notebook dedicated to main melody extraction.

This project aims to classify music based on structural features, specifically focusing on melodies independent from key or tempo. The primary focus is on Polish Christmas Carols (kolędy). The approach is based on three key references Serra et al. (2009), Chen et al. (2017), ChromaCoverId by @albincorreya that contribute to the understanding and methodology used in the project.

Problem Statement

The aim is similar to Cover Song Identification, which involves the classification of music based on its structural elements, particularly melodies. The goal is to distinguish between different versions of the same song performed by various artists while disregarding factors such as key or tempo variations.

Measures

The Harmonic Pitch Class Profile (HPCP) was employed as a reliable musical descriptor for Cover Song Identification. HPCP focuses on specific frequency ranges where the main musical information is found, reducing the influence of unwanted noise in the analysis. Additionally, HPCP takes into account harmonics (multiples of the fundamental frequency) to make the analysis independent of the specific tuning of the instrument. This helps in extracting features related to the musical notes' distinct colors or "chroma."

The project utilizes the dMax measure to compute distances between two samples represented by HPCP vectors.

The k-Nearest Neighbors (KNN) algorithm for effective classification was used, utilizing dMax as distance between datapoints.

Results

The project has yielded following results based on the two approaches:

1. Model Trained on Melody Extracts:

   • Accuracy: ~73%
   • The model performs moderately well and exhibits lower accuracy compared to the model trained on original versions of the Christmas carols.

2. Model Trained on Original Versions:

   • Accuracy: ~89%
   • This model outperforms the melody-only approach, emphasizing the importance of considering the complete musical structure.

Challenges and Insights

1. Generalization to New Data:

   • The model, particularly the one trained on melody extracts, does not generalize well to new data. Performance diminishes when applied to songs performed by artists not included in the training dataset.
   • Chroma feature extraction may be influenced by additional factors such as music genre, accompaniment, or voice timbre.

2. Overfitting:

   • While achieving high accuracy for various interpretations of Polish Christmas carols, the model seems overfitted.
   • Poor predictions for samples performed by artists outside the training dataset suggest a need for robustness against diverse musical styles.