Deep Learning Approach for Predicting Hit Potential of Music Tracks: A Comprehensive Analysis of Audio Features

Neeraj Kumar, Aiden Samuel, Amit Verma, Viyom Mittal

The primary objective of this research is to construct a deep learning model capable of predicting the success of a music track. The entire process consists of two models of regression and classification which utilize attributes related to audio signals. The purpose of this approach is to provide an accurate prediction of whether a song will become a hit or not, hence assisting musicians, producers, and record labels in making informed decisions. This research leverages advanced deep learning techniques for feature extraction and classification. The process begins with extracting audio features using VGGish, and processing embeddings using TensorFlow Hub. It also employs TensorFlow operations for feature extraction, specifically feeds audio waveforms into the model to obtain embeddings representing audio features such as timbre, pitch, and harmonics. These embeddings are then processed further for subsequent analysis. Next the process predicts the values for each feature using the corresponding model. After inverse transforming the predictions to the original scale, it calculates the mean prediction for each feature and stores them in a dictionary. Finally, the process concludes with the XGBoost classifier model that classifies them as hit or flop. The model demonstrated a high accuracy of 85% in predicting the hit potential of a track. Key audio features such as loudness and acousticness proved to be significant predictors, contributing to the majority of the model's decisions. The findings of this research have significant implications for the music industry. By accurately predicting a song's hit potential, musicians, producers, and record labels can make more informed decisions about the production and promotion of music. This could lead to reduced financial risk and increased success rates. Additionally, streaming platforms could use this model to recommend songs to users, enhancing their listening experience and engagement. Furthermore, these findings underline the importance of audio features in a song's success, encouraging a focus on those elements during the songwriting and production process.