Диплом (Master), University of Applied Sciences Upper Austria, Hagenberg, 2011. — 62 p.In this thesis I write about ways to compare audio in order to automatically find different versions of a song. A “version” is every new performance of a previously released piece of music. Most people can recognize if a given song is a version of another one. The same task performed by a computer although is a much more advanced problem involving many different aspects of music similarity. First, an overview of different musical features that are most likely to stay the same across versions is presented. Also ways how to extract these features and compare them to estimate the similarity of two songs. Different state of the art approaches for music version detection are discussed. Based on current algorithms, I implemented my own music version detection system. Therefore I extract beat–aligned harmonic pitch class profile features (HPCPs) out of a songs raw audio signal. To compare the features I used cross-correlation and dynamic time warping. All the processing steps that are necessary to generate HPCP features are described in detail. The presented version detection system is compared to another approach using Mel Frequency Cepstral Coefficients (MFCC). The results show that the presented approach is basically working but its performance is not better than the one of the other algorithm. Possible further improvements to the algorithm are discussed and the influence of different settings are presented. Also relevant problems that appeared while implementing the algorithm are mentioned. Further work directions are seen in exploring the settings involved in creating HPCP vectors to see if there are any ways to tune the features to work better for a specific genre.Introduction Music Version Detection – State of the Art The Harmonic Pitch Class Profile Evaluation and Discussion ConclusionThe covers80 Dataset The my30 Dataset
Чтобы скачать этот файл зарегистрируйтесь и/или войдите на сайт используя форму сверху.