Human Vocal Event Detection for Realistic Applications
Abstract: Human vocal event detection is very important in many areas including smart homes, smart cars, health-care, surveillance, behavior monitoring, etc. However, use of vocal event detection in realistic applications is challenging due to issues like variable speaker to microphone distances, difficulty to annotate vocal events like mental disorders, limitation of real audio data in realistic environments, difficulty to perceive event information from prosody or semantics of speech, etc. Audio analytic systems can usually be split into three parts, namely feature extraction and selection, feature modeling, and classification. In this research we propose novel solutions in each of these audio analytic parts to address the challenges to detect human vocal events in real applications adapting various machine learning, text mining and signal processing techniques. We hypothesize such solution will aid in developing and use human vocal event in smart homes, cars, health-care and surveillance applications. This proposal provides an overview of scope of human vocal event detection research, key challenges in real applications, hypothesized contributions, potential solutions, an evaluation plan and some preliminary results.
Committee Members: John A. Stankovic (Advisor), Hongning Wang (Chair), Yanjun Qi, Yuan Tian, John Lach