Online acoustic scene analysis based on nonparametric Bayesian model
Keisuke Imoto, Nobutaka Ono · 2016
In this paper, we propose a novel online method for analyzing acoustic scenes from sequentially obtained sounds. One prospective method for analyzing acoustic scenes is the use of a generative model of acoustic topics and event sequences in observed sounds, where the acoustic topic represents the latent structure of acoustic events associating an acoustic scene and acoustic events. This generative model is called an acoustic topic model (ATM). However, the conventional ATM employs a batch technique for estimating model parameters and cannot model sequentially obtained acoustic event sequences. Moreover, the number of classes of acoustic topics that lies in acoustic event sequences needs to be predetermined before observing acoustic events. However, the necessary number of acoustic topics for representing acoustic scenes varies in accordance with their contents, and this causes a mismatch between the actual number of classes of acoustic topics and the predetermined number of classes. In our method, the number of classes of acoustic topics can be automatically inferred from sequentially obtained acoustic event sequences on the basis of the online and nonparametric Bayesian technique. The experimental results of online acoustic scene estimation using real-life sounds indicated that the proposed method performed of acoustic scene classification better than the conventional ATM. In addition, the proposed method produced an efficient computation performance.