Prediction of Potential Hit Song and Musical Genre Using Artificial Neural Networks

Abstract

Accurately quantifying the goodness of music based on the seemingly subjective taste of the public is a multi-million industry. Recording companies can make sound decisions on which songs or artists to prioritize if accurate forecasting is achieved. We extract 56 single-valued musical features (e.g. pitch and tempo) from 380 Original Pilipino Music (OPM) songs (190 are hit songs) released from 2004 to 2006. Based on an effect size criterion which measures a variable's discriminating power, the 20 highest ranked features are fed to a classifier tasked to predict hit songs. We show that regardless of musical genre, a trained feed-forward neural network (NN) can predict potential hit songs with an average accuracy of Φ NN = 81%. The accuracy is about +20% higher than those of standard classifiers such as linear discriminant analysis (LDA, Φ LDA = 61%) and classification and regression trees (CART, Φ CART = 57%). Both LDA and CART are above the proportional chance criterion (PCC, Φ PCC = 50%) but are slightly below the suggested acceptable classifier requirement of 1.25*Φ PCC = 63%. Utilizing a similar procedure, we demonstrate that different genres (ballad, alternative rock or rock) of OPM songs can be automatically classified with near perfect accuracy using LDA or NN but only around 77% using CART.

Statistical Feature-Based Craquelure Classification

“…We use the standard gradient descent method in training a three-layer NN with linear activation function in the input nodes, and f ðxÞ ¼ 1:7159 tanh(2/3 x) in the hidden and output nodes closely following the procedure of our previous works. 11,15,16 The mentioned architecture class and training algorithms have been successful in characterizing various complex systems problems, from hit songs prediction 11 to public opinion forecast 15 to signal classification. 16 After extensibly playing with the networks free parameters such as learning rate, hidden nodes, and various activation functions, we have not seen a marked improvement in the test set accuracy of NN as compared to the LDA procedure.…”

“…11,12,14 That is, increase in the number of variables used for discrimination does not necessarily translate to improvement in accuracy. This effect referred to as \curse of dimensionality" has been recently shown to be critical in the accuracy of both linear and nonlinear classifiers that are used for potential hit songs prediction 11 and author attribution classification. 12 Here, we describe an effect size criterion-based technique that allows us to systematically choose the more relevant variable combinations for prediction.…”

“…These features are then ranked according to effect size ratio, 11,12 a procedure that we have shown recently to give a good discriminant variable combinations for accurate classification. We use four distinct classes of craquelure patterns, namely: 14th and 15th century Italian panels, 15th and 16th century Flemish panels, 17th century Dutch paintings on canvas, and 18th century French paintings on canvas.…”

Key-finding by artificial neural networks that learn about key profiles.

“…The notion of style or genre is too vague to be formalized in a fashion suitable for a classical rule governed system (Loy, 1991). However, neural networks are up to the task, and can classify musical patterns as belonging to the early works of Mozart (Gjerdingen, 1990), can classify selections as belonging to different genres of Western music (Mostafa & Billor, 2009), can evaluate the affective aesthetics of a melody (Cangelosi, 2010; Coutinho & Cangelosi, 2009; Katz, 1995), and that can even predict the possibility that a particular song has “hit potential” (Monterola, Abundo, Tugaff, & Venturina, 2009).…”

Bangla song genre recognition using artificial neural network

“…Liang et al [5] try to classify music ❒ ISSN: 2252-8938 genres using different techniques such as hidden Markov model (HMM's) and using canonical correlation analysis (CCA). In another paper, try to classify potential hit song and music genres [6]. For classifying those, they extract one dimensional audio features and using those they apply on different classification algorithms but feed-forward neural network gives almost 81% accuracy.…”