Recognition of Musical Instrument Using Deep Learning Techniques

Theoretical Background

Music instrument recognition aims to recognize the instruments played in the music clips. Humans can easily recognize the instruments based on the music characteristics such as pitch, loudness and the playing style. Realizing a system to recognize the instrument from a music clip is a difficult task due to the challenge in the extraction of essential information. According to Giannoulis & Klapuri (2013), timbre information is a major acoustic characteristic which enables human to recognize the sound sources. It is a vital feature to convey the uniqueness in musical instruments. However, the timbre information of a sound cannot be associated with any physical quantity (Agostini, Longari & Pollastri,2001). An exhaustive study made by Hall & Bahoura (2012) evidenced that MFCC and LPCC are a powerful and salient contribution for timbre characterization. They have also demonstrated the high score of instrument recognition using MFCC features. MFCCs are extensively studied and proved as reliable and robust features in speech and speaker recognition tasks. (Al-Ali, Dean, Senadji, Chandran, & Naik, 2017). Efficiency of MFCC is also experimented and evidenced in many of the MIR problems in particular genre classification, artist recognition, and music emotion recognition (Nalini & Palanivel, 2016). Chandwadkar & Sutaone (2013) studied the impact of feature selection and classification techniques for instrument identification. They worked with MFCC along with spectral features and autocorrelation coefficients to identify four instruments using various classifiers such as Decision Tree, K nearest Neighbours, Multi-Layer Perceptron, Sequential Minimal Optimization and meta-classifier. They also demonstrated the efficiency of MFCC features over the spectral features in instrument identification.