다중 특징 벡터를 이용한 고속 오디오 검색

Abstract

The types of information are changed text-based into various multimedia data such as speech, image, and moving picture. Therefore, it is necessary to study about searching algorithm. Previous keyword-based retrieval is not optimal for searching the multimedia data. Therefore, the studying is focus on the content-based retrieval (etc. MPEG-7) has been attracted. This thesis concentrated on the content-based retrieval and proposed a quick search method.

In the Audio Information Retrieval (AIR) System, it is important to extract feature vectors. Feature extraction is the process of computing a numerical representation that can be used to characterize a segment of audio. In this thesis, we use the features based on the Short Time Fourier Transform (STFT) and the zero-crossing rates. Firstly, Features based on the STFT are very common and have the advantage of fast calculation based on the Fast Fourier Transform algorithm. The STFT features can be classified into the spectral centroid, the spectral roll-off and the spectral flux. In the second place, the zero-crossing features have been used in the previous papers because of reducing the computation.

This thesis also proposes a new search using the preprocessing and code matching. The previous papers propose a time-series search method using the upper bound proof. It is assumed that similarity between the test and reference template shows considerable correlation from one time step to the next. Because the search algorithm using the upper bound proof computes upper bound on the similarity measures, this method can make possible the quick search. However the search speed of a time-series search method is very low at real time. Therefore this thesis proposes a method using the preprocessing to make up for this defect. Furthermore, we use the code matching method to reduce the matching rates.

This thesis is organized as follows : Section 2 overviews the previous time-series search algorithm. Section 3 explains the core part of our new algorithm and the new optimal combination of multiple features. Section 4 evaluates the accuracy and speed of the algorithm using multiple features. Finally Section 5 gives conclusions and future works.