Polyphonic Instrument Emotion Recognition using Stacked Auto Encoders: A Dimensionality Reduction Approach

Abstract

Instrument Emotion Recognition (IER) is the process of recognizing the emotion from instrumental music clips. The objective of this work is emotion recognition from polyphonic instrumental music. A dataset of instrumental music clips playing various instruments is collected for emotions neutral, happy, sad and fear. Mel Frequency Cepstral Coefficients (MFCC) and Chroma Energy Normalized Statistics (CENS) features are extracted from the instrumental music clips. The dimensionality reduction of extracted features is performed using Stacked Auto Encoder (SAE), and the instrument emotion recognition performance is analyzed. The performance of instrument emotion recognition was evaluated using recognition rate and equal error rate. A recognition rate of 74.6% and 62.1% and an equal error rate of 25.5% and 37.9% are achieved using SAE with MFCC and CENS features, respectively. The experimental results also show that the training and testing time of instrument emotion recognition is reduced with the decrease in the number of neurons in the code layer of stacked autoencoders.