Title1 : FxLMS-Based Active Noise Control: A Quick Review
Time1: Oct. 24, 2011 10：00-12:00 AM
Title2 : Speaker Recognition Feature Domain Investigation
Time2: Oct. 25, 2011 10：00-12:00 AM
Speaker: Dr. Waleed H. Abdulla
Place: Room1-415 FIT Building
Organizer: Research Institute of Information Technology (RIIT), Tsinghua University
The traditional approach to acoustic noise control uses passive silencers to attenuate unwanted sound waves. These silencers are relatively bulky, costly and ineffective for low frequency noise. To overcome these problems, Active Noise Control (ANC), in which an electro-acoustic system is responsible to create a local silence zone, has received considerable interest. ANC based controller calculates an antinoise signal with equal magnitude but opposite phase of the audible noise. This electrical antinoise signal is converted to an acoustic signal by a canceling loudspeaker placed close to the quite zone. The acoustic antinoise signal travels through the air and combines acoustically with the environmental noise.
By the end of the 1950's, several analog ANC devices were invented. However, all of analog ANC devices are not able to adapt to changing characteristics of the noise to be canceled and changing environmental conditions. This is because they cannot be controlled in an adaptive way. Only with the advent of digital technology the realization of adaptive ANC systems become possible. The theory of adaptive ANC, in which an adaptation algorithm automatically adjusts the ANC device, was established by Widrow in 1975; however the most significant progresses on this subject was reported in the recent two decades. The main challenge in adaptive ANC is the estimation of environmental noise with a short response time.
In this talk I present a short review on active noise control (ANC) with the emphasis on ANC systems implemented by using Filtered-x Least Mean Square (FxLMS) adaptation algorithm. The physical mechanism behind active noise control, based on which local silence zones can be created is detailed. Basic configurations for realization of ANC systems are then introduced. It is shown that FxLMS algorithm has been widely used in different types of ANC systems. Available theoretical work on analysis of FxLMS-based ANC systems is reviewed. Shortcomings of available theoretical findings are discussed.
Finally, recent advances in theoretical analysis of FxLMS-based ANC systems are introduced. These advances can be considered as the recent contributions made by the speaker and his group.
Speech is the most natural and easy way of communication and it is a rich information carrying signal. This signal carries information about speaker’s identity, emotion, age, gender, and many more. Our research focuses on extracting the identity information and investigating ways of introducing robust information carrying features in noisy environments.
In this seminar we spotlight on mining speech signals to extract information about speakers’ identity. The process of identification comprises two stages; feature extraction and classification. In feature extraction, distinctive attributes from human voices are extracted to refer to the speakers. In classification, the extracted features are processed to find the identity of the most probable speaker. So far, not a single technique can be claimed to offer human-like identification performance in all environments (clean and noisy). We strongly believe that the major problem is in the feature extraction stage rather than in the classification stage.
Several features will be discussed, such as the perceptual log area ratio (PLAR) feature and the Gammatone based correlation ICA features. The performances of the investigated features as compared to the commonly used features will be discussed.
Throughout this talk, we will try to put a case to tell that there is still room to develop more advanced features than the commonly used ones to serve the human recognition objective.