REAL-TIME BEAT TRACKING SYSTEM FOR MUSICAL AUDIO SIGNALS BASED ON

• Main Author: Seandy Sunjoko-NIM : 13502038, Yohanes
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/9675
• Institution: Institut Teknologi Bandung

Abstract: <br /> <br /> <br /> <br /> <br /> <br /> In this final project has been implemented a real-time beat tracking system for musical audio signals named Rebana. While the song is being played, the system makes predictions about temporal position, tempo, and beat type of the next beat. <br /> <br /> <br /> <br /> <br /> <br /> There are several problems on real-time beat tracking. First, beats are not explicitly expressed in musical audio signals, but may be tracked by inferring various musical elements extracted from the signals, e.g onset times and chord changes. Second, real-time beat tracking faces a problem, i.e gardenpath error, thus multiple predictions are necessary. In order to make multiple predictions, in this final project a multiagent system is used. <br /> <br /> <br /> <br /> <br /> <br /> As information about onset times and chord changes is extracted from the received signals, agents keep making and evaluating predictions according to their own strategy. The predictions are then grouped and are selected by an entity called the manager based on some criteria. This system has been successfully implemented on Mandrake Linux 9.2 in a PC with single processor. <br /> <br /> <br /> <br /> <br /> <br /> It makes use several programming tools, i.e GCC, PortAudio, FFTW, and GTK+. Multithreading is applied in the implementation because the system is urged to perform several tasks together at a time. <br /> <br /> <br /> <br /> <br /> <br /> Its performance is tested by considering three sides, i.e the technicals, the correctness of selected predictions, and the response time. From the technical side, the tests reveal some cases that require special treatments related to the agents work. From the side of correctness of selected predictions, the test result of the system is quite good. It is shown by the high average of correctness value among 17 times of tracking, i.e 62.65%. The most common failures which often occur are that the temporal position of the predictions lies between two <br /> <br /> <br /> <br /> <br /> <br /> successive correct beats and the presence of unstable beats yielded by the manager. From the side of response time, the tests show that response time generally exceeds inter-beat interval as the upper limit of acceptable response time. It is obviously shown as there are only three of 17 times of tracking giving tolerable response time. The factors causing it are the time that must be repeatedly spent by the manager to wait as predictions from the agents are gathered, the time spent in checking <br /> <br /> <br /> <br /> <br /> <br /> chord changes, and the scheduling mechanism of threads performed by the operating system. <br /> <br /> <br /> <br /> <br /> <br /> Therefore, this beat tracking system is unfortunately too slow to be acknowledged as a real-time application. <br /> <br />