We propose a search method for detecting a query audio signal fragment on long audio recordings. The query signal is assumed to be captured by a portable terminal in the real world. A major problem in this kind of searching is that the features of query sound may include distortions due to terminal characteristics or environment noises. The method proposed here comprises independent normalization and robust subspace spanning. The former is used to absorb additive noise and frequency characteristics. The latter is used to choose frequency bands that minimizes the effect of feature distortions. The experiments using audio signals received in the real world prove the effects of the proposed method; for example, the search accuracy was 84.29 when a 13 hour audio recording was searched through.
展开▼