잡음 환경에 강건한 음향 이벤트 위치 추정 및 탐지에 대한 연구

Abstract

SELD(sound event localization and detection) aims to classify sound events, detect their onset and offset, and estimate their direction. Recent advancements in deep learning have led to the proposal of various SELD methods based on deep learning techniques. However, existing deep learning-based SELD methods have mostly been developed and evaluated on audio data with minimal noise. Consequently, applying these methods in environments with high levels of noise often yields suboptimal performance. To address this issue, some approaches preprocess the data by applying noise reduction techniques before performing SELD, but this can be time-consuming. To overcome these challenges, this study proposes a SELD model that incorporates deep learning-based speech enhancement techniques. The proposed model combines a U-net architecture for speech enhancement and a SELDnet for SELD, enabling end-to-end processing of noise reduction and SELD tasks. To evaluate the proposed method, experiments were conducted using datasets containing diverse levels of noise. The results demonstrated that the proposed model outperformed existing SELD models in environments with high levels of noise, although it showed slightly lower performance in noise-free environments.