Sound event detection method capable of analyzing any environmental sound

2021 Fiscal Year Final Research Report

• Project/Area Number: 19K20304
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 61010:Perceptual information processing-related
• Research Institution: Doshisha University (2020-2021); Ritsumeikan University (2019)
• Principal Investigator: Imoto Keisuke 同志社大学, 理工学部, 准教授 (90802116)
• Project Period (FY): 2019-04-01 – 2022-03-31
• Keywords: 環境音分析 / 音響イベント検出 / 音響シーン分類

Outline of Final Research Achievements

This research project aims to develop a method to achieve reasonable performance in environmental sound analysis, which is one of the most important topics in acoustic processing, even when only a small amount of environmental sound data is available. In particular, we investigated a method for sound event detection (SED) based on the co-occurrence of environmental sounds and the omnipresence of sound events in an acoustic scene. During the research period, we have proposed a method considering the co-occurrence of acoustic events with deep learning methods, a multi-task learning method of SED and acoustic scene classification, and a model learning technique that does not cause performance degradation even when there is a data imbalance between sound event classes, showing that sound events can be detected with high accuracy.

Free Research Field

音響信号処理

Academic Significance and Societal Importance of the Research Achievements

人間の耳のように，様々な環境音の種類を聞き分ける技術が実現できれば，補聴器などの聴覚補助システムのみならず，公共スペースでの自動監視システム，高齢者や乳幼児の見守りシステム，自動運転の補助，環境の自動モニタリング，知的ロボットなど様々なサービスに広く貢献できる．このように，環境音分析は音響処理の中でも非常に重要な技術として位置づけられる．また，画像/動画の分析などの技術と組み合わせることで，人間の知覚を模した人工知能を実現することも可能となる．