This paper proposes a search method for music signals used as back ground music (BGM) in a long audio signal (stored signal). In such a search, a music signal is given as a reference signal and every segment including that music signal as BGM must be detected from the stored signal. The segments, even when they include the reference signal, differ considerably from the reference signal due to the presence of other audio signals such as voices. Therefore, it is difficult to detect segments that include the reference signal by direct matching of the reference signal and the stored signal. To solve this problem, we first chose small area spectra decomposed from the spectrum of the reference signal, and we search for similar segments for each decomposed spectrum in the stored signal. Then, we integrate the local similarities of detected small segments and find the segments including the reference signal. Searching for only the small segment with larger local similarity than the threshold greatly accelerates the search. In the experiments, a 15-second reference signal was detected from the stored signal in about 8 seconds by utilizing Time Series Active Search as the searching process for small area spectra.
展开▼