A smart eavesdropping system : speech enhancement for law enforcement agents

• Main Author: Yeo, Nadia Xue Min
• Other Authors: Khong Andy Wai Hoong
• Format: Final Year Project
• Language: English
• Published: 2015
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/64711
• Institution: Nanyang Technological University

Law enforcement agencies always face the challenge of being able to successfully eavesdrop on a targeted subject clearly in a noisy restaurant setting. To solve this problem, it is important to develop a smart eavesdropping system that is able to separate and clearly hear the conversation of the desired subject from the noisy background. The purpose of the project is to test on an existing system and integrate additional functions that further enhance the user’s experience at the back end. The concept of beamforming is applied to listen on the targeted subject in a noisy environment. The student would study the beamforming techniques, namely delay-and-sum beamforming and Minimum Variance Distortionless Response (MVDR) prior to testing out on the existing system under a virtual environment. Subsequently, the student will examine the available face detection methods before deciding on the most appropriate approach (Viola-Jones Face Detection Framework) to be developed. This face detection and tracking function will be integrated into the existing live video function of the Graphic User Interface (GUI) as an additional feature. After performing tests on the simulated system, the student discovered that the voice of the targeted subject could be heard more clearly using the MVDR method. Upon successfully developing the face detection and tracking function, the student has tested out and obtained relatively high success rates of face detection and tracking. It is concluded that MVDR is the better approach in filtering out background noises and further enhancing the voice quality of the targeted subject. For the face detection and tracking function, the Viola-Jones Face Detection Framework is the better approach as it enables fast computation of the detected faces; low latency in live video rendering. However, the face detection and tracking function can be further improved, as the Viola-Jones algorithm is unable to detect faces tilted to certain angle. Additional point tracking function (KLT algorithm) can be augmented to ensure higher face detection rate.