Automatic Sri Lankan Traditional Musical Instruments Recognition In Soundtracks

Abstract

Musical instrument recognition is an essential aspect of music information retrieval, and nowadays, audio signal processing is an active research domain. Automatic identification of traditional music instruments from the soundtracks is one of the applications which combines signal processing and machine learning techniques. So, this paper presents an application to automatically recognise the Sri Lankan traditional music instruments from long music tracks. Soundtracks of Ten (10) instruments were collected from various domain experts to demonstrate the proposed method. Four different features are extracted and compared from collected soundtracks to find the most suitable feature for Sri Lankan traditional musical instrument sounds. Using Principal Component Analysis (PCA), the six (06) most significant features were selected from twenty (20) Mel Frequency Cepstral Coefficients (MFCC) features. Then two (02) machine learning algorithms (K-NN, SVM) are used to classify the traditional instruments’ soundtracks separately and classified. By outperforming other models, the SVM model with MFCC features provided 86.8% of the highest accuracy.