Real-Time Hit Classification in a Smart Cajón

Frontiers in ICT

Turchet, Luca¹  McPherson, Andrew¹  Barthet, Mathieu¹ 

[1] Centre for Digital Music, School of Electronic Engineering and Computer Science, Queen Mary University of London, United Kingdom

关键词: smart musical instruments;    Internet of Musical Things;    music information retrieval;    machine learning;    Sensor Fusion;    cajó;    n;   

DOI  :  10.3389/fict.2018.00016

学科分类：计算机网络和通讯

来源: Frontiers

【 摘 要 】

Smart musical instruments are a class of IoT devices for music making, which encompass embedded intelligence as well as wireless connectivity. In previous work, we established design requirements for a novel smart musical instrument, a smart cajón, following a user-centred approach. This paper describes the implementation and technical evaluation of the designed component of the smart cajón related to hit classification and repurposing. A conventional acoustic cajón was enhanced with sensors to classify position of the hit and the gesture that produced it. The instrument was equipped with five piezo pickups attached to the internal panels and a condenser microphone located inside. The developed sound engine leveraged digital signal processing, sensor fusion, and machine learning techniques to classify the position, dynamics, and timbre of each hit. The techniques were devised and implemented to achieve low latency between action and the electronically-generated sounds, as well as keep computational efficiency high. The system was tuned to classify two main cajón playing techniques at different locations and we conducted evaluations using over 2000 hits performed by two professional players. We first assessed the classification performance when training and testing data related to recordings from the same player. In this configuration, classification accuracies of 100% were obtained for hit detection and location. Accuracies of over 90% were obtained when classifying timbres produced by the two playing techniques. We then assessed the classifier in a cross-player configuration (training and testing were performed using recordings from different players). Results indicated that while hit location scales relatively well across different players, gesture identification requires that the involved classifiers are trained specifically for each musician.

【 授权许可】

CC BY