Abstract

This model presents a content-based multilingual song recommendation system developed using unsupervised machine learning techniques.Designed to support five languages—Hindi ,Tamil ,Korean, English ,and a miscellaneous type—the system takes a user’s input of a language and a specific song title, and identifies similar songs within that language to ensure linguistic and cultural relevance. To build the recommendation engine, we curated a dataset of 61,711 songs across the supported languages using Kaggle as our source. Audio features such as tempo, valence, energy, acousticness, and danceability were extracted and standardised. We then applied Kernel Principal Component Analysis (Kernel PCA) with an RBF kernel to reduce dimensionality while preserving non-linear feature relationships. The reduced feature set was clustered using K-Means, which produced more distinct and meaningful groupings than DBSCAN. K-Means was fine-tuned by using K-Means++ for centroid initialization and the Elbow Method by analyzing the Within-Cluster Sum of Squares (WCSS) to determine the optimal number of clusters per language, while the best epsilon values were used for DBSCAN for comparison.Once clusters were formed, cosine similarity was employed within the relevant cluster to retrieve songs most similar to the user’s input track. This hybrid approach ensures that recommendations are both musically coherent and computationally efficient. The system consistently delivered results aligned with user preferences in mood, genre, and rhythm. Future improvements may include integrating lyrics-based features, deep learning models for better embeddings, and user feedback mechanisms to enhance personalization further.

Keywords: Multilingual+Content-Based Song Recommendation, Unsupervised Learning, Clustering, KMeans, DBSCAN

In a vast multilingual music ecosystem, users often seek songs that match the mood, rhythm, or energy of a known track, but struggle to find similar songs, especially across different languages. Our system addresses this by using audio feature based clustering and similarity metrics to recommend songs closely aligned with the user’s musical preferences.We avoid using other user's playlist or listening history(behavioural data ) to determine songs for recommendations. Given a track name and language, the model identifies a group of similar songs using K-Means clustering and cosine similarity, helping users discover songs that sound similar and match the language they prefer.

We also tried using DBSCAN for clustering our data after reducing it using Kernel PCA. DBSCAN is useful because it can find clusters of any shape and doesn’t need us to decide the number of clusters in advance. But it didn’t work well with our dataset. When used on 2D-reduced data, DBSCAN marked most songs as outliers and couldn't form proper clusters. Even after finding the best value for its settings using the Elbow Method and applying it to 4D-reduced data, it still gave very few or unclear clusters. This happened because our song data is very dense and smooth, and DBSCAN depends on finding gaps in the data to work well. In the end, K-Means gave us much better and clearer clusters, so we chose it for our final model.

Clusters of different Languages Models

Elbow Plot for Hindi Language Model

Clusters of Hindi Songs after dimensionality reduction using PCA

Song Recommendation on WebPage for "Shape of You"