Fonara Embedding Model 🎼

Aim 🎯

The Fonara Embedding model was created to produce embeddings for Greek songs and compare their similarity, allowing similar-sounding Greek songs to be identified using Cosine Similarity.

Base Model 🇬🇷

The CultureMERT-95M model was finetuned on a custom dataset of Greek song snippets (which cannot be redistributed due to copyright).

Training Data Details 📖

A small curated dataset of approximately 3.6k songs was used to finetune the model, including songs from many different Greek genres such as rebetika, laika, nisiotika, etc. The aim of this was to create a more diverse training set to drive the model towards generalization across different genres, rather than memorization of particular songs.

Metrics 📐

Preliminary benchmarks were performed on a held-out test set; however, extensive testing to validate the usefulness of the model outside of context of the Fonara app has not been conducted.

Training Details and Hardware 🖥️

A single T4 GPU was used to finetune the base model over 8 epochs, provided generously for free by Kaggle.

Usage

For details on the use of the model and its application, please visit the GitHub repo.