Tone Quality Recognition of Instruments Based on Multi-feature Fusion of Music Signal

The traditional expert-based instrumental music evaluation strategy can’t meet the requirements of the rapidly accumulated audio data. The traditional strategy not only takes a high cost of human’s energy and time but also may have some problems on consistency and fairness of judgment. This paper aims at designing a complete recognition and evaluation strategy to automatically identify the timber of wind instruments. We take the clarinet as example and propose a strategy based on multi-feature fusion and random forest. First, we use the identification of fundamental frequency algorithm to automatically distinguish the notes performed by the instruments. Second, we extract 3 types of features including MFCC, brightness and roughness to describe the instrumental signals. Then, considering two kinds of variants: note and tone quality, we design 5 strategies to remove the influence of different notes in the evaluation of tone quality. By analyzing these strategies, we explore the optimal strategy for the recognition. The final evaluation results over 840 music slices demonstrate the effectiveness of this method.