Noisy Text Analytics
Noisy text data is found in informal settings (online chat, SMS, e-mails, message boards, etc.) and in text produced through automated speech recognition or optical character recognition. Noise severely degrades the performance of other information processing algorithms (e.g. classification, clustering, summarization). In this work techniques to overcome the noise and still do reasonable analytics are being developed. The focus is on extracting business intelligence from contact center data (voice and text) [AND07][IJCAI 07][ACL-COLING06].
Biological Knowledge Discovery Infrastructure
Biological text data is getting generated at a very fast rate. New research findings are published and made available online (for example, PUBMED). As a medical researcher, drug designer, biologist, life scientist, chemist, ... this is the literature one reads to come up to date with the latest in the field. However, because of the volume of data it is very difficult to home in on the relevant information. In this project new methods for biological knowledge discovery are being developed [CIKM 03][IBM RD 04][IBM SJ 04].
Audio-Visual Speech Technologies
In this project worked on the twin problems of multimodal speech recognition and audio driven facial animation. The audio and video provide orthogonal information in many cases and their combination is been shown to aid in speech recognition greatly [MMSP 99][ICPR 00]. Also in this project methods were suggested for tracking of lips on a talking face and the extraction of visual features, for speech recognition, from the lip region [ICPR 00]. Also techniques for audio driven facial animation using morphing based [ICME 01]and other approaches [ICASSP 03][Viseme Patent] were suggested. Techniques for adapting the phone set of one language to another were developed [ICSLP00(1)] [ICSLP00(2)]. A translingual visual speech synthesis engine based on this research was developed [ICME 00][IEEE T-MM 04].
Underwater Image Processing
Worked as part of the team developing algorithms and software for the Department of Electronics sponsored work on ADvanced Object VIsualization Techniques (ADOVIT) for the underwater scenario from August 1993 to September 1996 [ADOVIT 94][ADOVIT 95]. This work involved forming visual images using sonar data. Sonar data is very sparse and noisy. A multiframe imaging Technique was developed to reduce speckle and noise [UUST 95]. Using the sonar data as hard constraints it is merged with a shape from shading model obtained of the scene from visual images. Segmentation and surface understanding techniques are then used to form a dense image for a human observer [UUST 95].
Coding for AWGN and Fading Channels
How Good is a code designed for the AWGN Channel? Tried to answer this question by obtaining a lower bound on the largest achievable rate vs Euclidean distance of the code. This work also suggests constellations in 2,3 and 4-dimensions over which asymptotically good codes may be found [IEEE T-IT 02][ISIT98][NCC 99]. Over Fading channels worked on TCM schemes with asymmetric PSK signal sets specially modeled to maximize performance [ISIT 98][IEEE T-VT 00].