Exploiting Temporal Feature Integration for Generalized Sound Recognition

摘要

This paper presents a methodology that incorporates temporal feature integration for automated generalized sound recognition. Such a system can be of great use to scene analysis and understanding based on the acoustic modality. The performance of three feature sets based on Mel filterbank, MPEG-7 audio protocol, and wavelet decomposition is assessed. Furthermore we explore the application of temporal integration using the following three different strategies: (a) short-term statistics, (b) spectral moments, and (c) autoregressive models. The experimental setup is thoroughly explained and based on the concurrent usage of professional sound effects collections. In this way we try to form a representative picture of the characteristics of ten sound classes. During the first phase of our implementation, the process of audio classification is achieved through statistical models (HMMs) while a fusion scheme that exploits the models constructed by various feature sets provided the highest average recognition rate. The proposed system not only uses diverse groups of sound parameters but also employs the advantages of temporal feature integration.