Acoustic event recognition using dominant spectral basis vectors

This paper proposes a novel filter bank composed of dominant Spectral Basis Vectors (SBVs) in a spectrogram. Spectral envelopes represented by the SBVs have shown to be excellent characteristic features for discriminating different acoustic events in noisy environment. Non-negative Matrix Factorization (NMF) and non-negative K-SVD (NKSVD) for part-based and holistic representations extract dominant SBVs from a spectrogram. The effectiveness of the proposed method is demonstrated on a database of real life recordings via experiments, and its robust performance is compared to conventional methods.