Seminar "Towards audio event detection in continuous streams"

Kính gửi Quý Thầy/Cô,
Phòng ĐTSĐH&KHCN trân trọng kính mời Quý Thầy/Cô tham gia buổi Seminar "Towards audio event detection in continuous streams"

Thời gian: 13 giờ 30 - 16 giờ 00 ngày 14/7/2017 (thứ 6)

Địa điểm: Phòng A116, Trường ĐH Công nghệ Thông tin

Người báo cáo: Phan Quốc Huy.

Topic: Towards audio event detection in continuous streams

Abstract: This talk will present a regression-based detection system for audio event detection on continuous streams. The idea is to encode the temporal structure for audio events with a regression model for event onset and offset estimation. The proposed framework is further shown its support on other important perspectives of a detection system, namely early detection, multi-channel fusion, and false positive reduction. The goal of early detection is to detect the target events early in time even when their partial duration are observed without losing the overall detection performance. Concerning multi-channel fusion, the presented approach offers a simple but efficient fusion framework on which channel-wise confidence detection scores are additively accumulated to form the fusion detection scores. Finally, the performance gap between audio event detection and classification is studied and leveraged for an enhanced detection pipeline in which a high-quality classifier is employed to verify and reject false positive ones outputted by a detection system.

Trân trọng kính mời Phòng ĐTSĐH&KHCN

File đính kèm: