Audio Classification Based on Weakly Labeled Data
Chieh-Feng Cheng, David V. Anderson, Mark A. Davenport, Abbas Rashidi · 2018
Audio event detection and classification are critical tasks in the analysis of multimedia data. Most current research on this topic focuses on processing strongly labeled data and using fully supervised machine learning techniques. However, many sources of multimedia data lack detailed annotation and rather have only high-level meta-data describing the main content of various long segments of the data. We propose a novel framework to perform audio classification when working with such weakly labeled data. A traditional approach to this problem is to use techniques for strongly labeled data and then to deal with the weak nature of the labels via post-processing. In contrast, our approach directly addresses the weakly labeled aspect of the data by classifying longer windows of data based on the clustering behavior of the acoustic features over time. We evaluate the proposed framework using both synthetic datasets and real data and demonstrate that our method can significantly outperform the traditional approach.