National Technical University of Athens (NTUA)
Πολυ-αισθητηριακή ακουστική ανάλυση περιβάλλοντος
Abstract
dc:descriptionIn our Dissertation we examine the problem of Acoustic Event Detection (AED) in multi-channel smart-space environments. AED constitutes a major part of the computational auditory analysis field, and its main goal is the automatic end-pointing and classification of each sound event present in an audio clip. In our research we focus on developing methods for exploiting the information from multiple microphones for detecting events under challenging and overlapping conditions. At first, we focus on the detection of human speech events in smart homes consisting of multiple rooms, equipped with multiple microphones. For this purpose, we develop a novel two-step room-localized Speech Activity Detection (SAD) system, appropriate for voice-enabled applications. In its first step, our system efficiently combines the signals from multiple microphones to produce temporal speech segmentation, and in the second step it extracts novel room-discriminant multi-channel features to locate the speaker at the room-level. Our system performs robustly and compares favorably to deep-learning based alternatives. Then, for the general AED task, we focus on the challenging overlapping scenario and experiment with non-negative matrix factorization (NMF)-based approaches. In the process, we investigate ways to improve the detection step of well-known NMF baselines, to increase the robustness of classifier-based NMF systems in highly overlapping conditions, and finally to develop an efficient multi-channel NMF system for detection tasks. Finally, we employ deep-learning methods for overlapping AED when the number of different event classes is large. In this direction, we propose the combination and joint training of a multi-channel sound source separation network with a multi-label AED network, achieving improved results over traditional neural network approaches. For our experiments, we employ several synthetic and real databases recorded in suitable multi-microphone smart-space environments.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2021
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Giannoulis, Panagiotis
- Γιαννούλης, Παναγιώτης
Subjects
dc:subject × 12- Εντοπισμός ακουστικών γεγονότων
- Πολυ-καναλική επεξεργασία ακουστικών σημάτων
- Ακουστικά γεγονότα με επικάλυψη
- Acoustic event detection
- Multi-channel signal processing
- Overlapping acoustic events
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Ρομποτική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Robotics
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/51375