Unsupervised detection of non-stationary segments based on single-basis non-negative matrix factorization for effective annotation

We propose novel methods for automatically detecting non-stationary segments using non-negative matrix factorization (NMF) with aiming to effective sound annotation. For acoustic event detection or acoustic scene analysis, preparing a sufficient amount of training data is important. However, listening to all recorded signals and annotating them are very time-consuming. Assuming that the observed acoustic signal consists of a single stationary background sound and various short acoustic events, we apply NMF with single spectral basis for the spectrogram of the observed signal. Since it is expected that stationary background sound is well represented by NMF, the residual divergence is a good measure to detect non-stationary sound. Through experiments on nine real-world recordings of outside environmental sounds, we show that our methods effectively detect a variety of non-stationary sound segments.