Audio-Based Retrieval of Musical Score Data


Deprecated: implode(): Passing glue string after array is deprecated. Swap the parameters in /var/www/html/arg/administrator/components/com_jresearch/helpers/publications.php on line 269

Deprecated: implode(): Passing glue string after array is deprecated. Swap the parameters in /var/www/html/arg/administrator/components/com_jresearch/helpers/publications.php on line 269

Deprecated: implode(): Passing glue string after array is deprecated. Swap the parameters in /var/www/html/arg/administrator/components/com_jresearch/helpers/publications.php on line 269

Deprecated: implode(): Passing glue string after array is deprecated. Swap the parameters in /var/www/html/arg/administrator/components/com_jresearch/helpers/publications.php on line 269
Subedi, Bishwa Prasad

Abstract

Given an audio query, such as polyphonic musical piece, this thesis address the problem of retrieving a matching (similar) musical score data from a collection of musical scores. There are different techniques for measuring similarity between any musical piece such as metadata based similarity measure, collaborative filtering and content-based similarity measure. In this thesis, we use the information in the digital music itself for similarity measures and this technique is known as content-based similarity measure. First we extract chroma features to represents musical segments. Chroma feature captures both melodic information and harmonic information and is robust to timbre variation. Tempo variation in the performance of a same song may cause dissimilarity between them. In order to address this issue we extract beat sequences and combine them with chroma features to obtain beat synchronous chroma features. Next, we use Dynamic Time Warping (DTW) algorithm. This algorithm first computes the DTW matrix between two feature sequences and calculates the cost of traversing from starting point to end point of the matrix. Minimum the cost value, more similar the musical segments are. The performance of DTW is improved by choosing suitable path constraints and path weight. Then, we implement LSH algorithm, which first indexes the data and then searches for a similar item. Processing time of LSH is shorter than that of DTW. For a smaller fragment of query audio, say 30 seconds, LSH outperformed DTW. Performance of LSH depends on the number of hash tables, number of projections per table and width of the projection. Both algorithms were applied in two types of data sets, RWC (where audio and midi are from the same source) and TUT (where audio and midi are from different sources). The contribution of this thesis is twofold. First we proposed a suitable feature representation of a musical segment for melodic similarity. And then we apply two different similarity measure algorithms and enhance their performances. This thesis work also includes development of mobile application capable of recording audio from surroundings and displaying its acoustic features in real time.

Research areas

Year:
2014