실제 잡음 환경에 강인한 오디오 핑거프린팅 기법

Abstract

Recently, content-based audio identification by an audio fingerprinting scheme that can retrieve audioinformation without any text-based query has been proposed and implemented. This paper offers a solution to this challenging task, using a music signal query to retrieve polyphonic music items by matching it to preindexed audio reference. In a real-environment, however, sound recordings are commonly distorted by channel and background noise, and the performance of audio identification is greatly degraded by them. This paper introduces the conventional audio fingerprinting schemes such as stochastic pattern modeling method and audio hashing method. We proposed a new distance metric to measure the similarity of two probability distributions in the stochastic pattern modeling. In addition, this paper improved the conventional audio hashing scheme using some modifications of both frequency filtering and temporal filtering for robust identification in real-world environments that are noisy and error-prone. As an alternative to temporal filtering, the RASTA filter, instead of a high-pass filter, is used to immune to channel distortion. Our experimental results show that the proposed methods improve channel and noise robustness in audio identificatoin.