Beat Tracking Based On Max-Min Distance Means

The rapid developments of the Internet and the advancements of the web technologies have made it possible for human beings to have access to huge amounts of on-line music data, such as music acoustic signals, lyrics, style or mood labels and the list of other users. People have more fun with the progress when they listen to the music. At the same time, an issue of how to organize this information has raised. It's becoming a popular field that how computer programs can assist human in their music experience. And it's also make contributions to the developments of MIR(music information retrieval). MIR is the interdisciplinary science of retrieving information from music. MIR is a small but growing field of research with many real-world applications. Those involved in MIR may have a background in musicology, psychology, academic music study, signal processing, machine learning or some combination of these.Beat tracking is a basic part of MIR research. For humans, tracking the beat is an almost natural task. One taps his foot or nods his head to the beat of the music. For the computers, beat tracking is to imitate human perception of music. The past 20 years, a large number of researches have been made in beat tracking field already, and more and more applications have connection with beat tracking.After deeply study of beat tracking related research, a novel beat tracking algorithm has been proposed, based on max-min distance means, combined with music basic theory and audio signal processing technology. The core of the algorithm has three parts: determination of the first beat, the BPM feature extraction and the peak picking. The innovation point is applying clustering algorithm in the beat tracking. The peak extraction problem is solved as a classification problem. This thesis includes the following aspects:First of all, pre-process the music signal to unify the sampling frequency and amplitude range. Process the 1-2s of the music in time domain. Through the analysis of the energy spectrum, the first beat can be located.Then, the audio data is transformed to the frequency domain via STFT. According to the perceptual properties of the human auditory system, a logarithmic calculation of spectral amplitude phase information, combined with the half wave rectifier output the onset strength curve and its peak. The BPM feature is extracted by using the autocorrelation function.Finally, the beats of the music signal are identified according to the relationship between music tempo and beat, together with effective clustering of onset strength curve's peaks.In this thesis, the experiment is carried out with MIREX2006 test data and compared with some algorithms in MIREX2013. The experimental results show that, the proposed algorithm can accurately and effectively detect the beat sequence of music signals with different styles, different music rhythm type. The four evaluation indicators of the algorithm P-score, Cemgil, CMLc and AMLt, have reached 57.35510, 38.70537, 17.15240 and 47.25912. It also turns out that the proposed algorithm has great advantages in the global correctness and continuous correct rate, and possesses better comprehensive performance.