Generalized Sound Recognition in Reverberant Environments

Computational Auditory Scene Analysis (CASA) is typically achieved by statistical models trained offline on available data. Their performance relies heavily on the assumption that the process generating the data along with the recording conditions are stationary over time. Nowadays, the focus of CASA is moving from structured, well-defined scenarios to unrestricted scenes with realistic characteristics where the stationarity assumption might not be true. Therefore, there is a high demand for methodologies and tools dealing with a series of problems tightly coupled with such non-stationary conditions, such as changes in the recording conditions, reverberant effects, etc. This paper formulates these obstacles under the concept drift framework and explores the two fundamental adaptation approaches: active and passive. The overall aim is to learn online the statistical properties of the evolving data distribution and incorporate them into the recognition mechanism for boosting its performance. The proposed CASA system encompasses: (a) a concept drift detector and (b) an online adaptation module. The proposed framework was evaluated in the auditory analysis of three environments (office, meeting room, and lecture hall) with diverse characteristics (dimensions, reverberation times, etc.) The corpus is based on a combination of professional sound event collections. We report encouraging experimental results in terms of classification rate, false positive rate, false negative rate, and detection delay.