An unsupervised auditory scene analysis system using incremental low-dimensional embedding

This paper addresses the low-dimensional embedding of sounds towards unsupervised auditory scene analysis. Summarizing long-time recordings by mapping them into lowdimensional space is an essential task in long-time environmental monitoring. In this paper, we propose a novel low-dimensional embedding system using an incremental embedding algorithm. To analyze long-time recordings, we design an incremental system consisting of recording, feature extraction, low-dimensional embedding, and visualization. Recently, many low-dimensional embedding methods for acoustic scenes have been studied; however, applicability of these methods to the long-time recording is not adequately evaluated. Thus, this paper describes the construction of the scene analysis system and evaluates the performance of this system. In this paper, we especially focus on two important viewpoints in long-time monitoring: incremental methods and effects of noisy data. To realize an incremental system, we use Self-Organizing Nebulous Growths (SONG), which can incrementally construct a low-dimensional embedding space. Also, in our experiments, we apply our system to bird song analysis under noise conditions. By the preliminary experiments using benchmark datasets, we discover noise sensitivity of our system and applicability to environmental monitoring.