8    Conclusions

Average Mutual Information can be used as a method for beat tracking.  One of the main problems addressed in any research related to beat detection is the quantitative evaluation of tracking accuracy, in particular for music files.  In this study and in most of the studies reviewed, human interaction is required for the measure of performance in a beat tracking system such as counting the number of beats in a period of time or labeling the beat positions in an audio file.  Most of the studies in beat detection measure the beat tracking accuracy using their real-time systems over 40 or 50 songs of different genres [12], [31], and determine if the system performed well or not.  In this study several types of audio files were tested to determine if AMI could be applicable in solving the beat detection problem.

The system performed better with decimation factor N=2, maxtau=2, bin=64, fs=8 kHz and length of 5 sec.  The algorithm of detection is very simple and if it is implemented in C++ the response times could be very suitable for real-time implementations.  In addition, if a proprietary function of AMI were implemented just for beat detection instead of using the one provided by TSTOOLS, the program could run faster and probably yield better results.  The next stage of this research is the implementation of a real-time system that could use the method presented (which detects the beat rate), but the way to determine the beat phase or when exactly the next beat is going to happen, needs to be defined.

The system also presented some capacity to detect quality in the beats as the quarter note or the strong beat (half note).  If this algorithm is combined with others, which handle hypothesis agents as the Goto and Muraoka [12], it could generate some interesting improvements.  Also AMI should be tried in beat detection algorithms that already use autocorrelation in some of their states, as Goto and Muraoka [12], because audio signals are more non-linear.

Further experiments could be achieved by adding algorithms capable of making decisions involving the beat detection values yielded by the AMI.  More intelligent algorithms are required to interpret AMI results.  But even human beings, whose own innate rhythm recognition skills could have subjective appreciation and different perception depending if they are a trained musician or just a novice, are prone to error in accomplishing this difficult perceptual task.