Chanwoo Kim

Major Graduate Projects

• DARPA GALE Project: Robust front end for the DARPA GALE project. Applied a new feature set that we developed for the 180-hr training set and the large vocabulary Broadcast News test set.
• NSF Auditory Signal Processing Project: Developed features motivated by auditory processing, binaural processing, and poly-aural processing.
• Samsung Speech-to-Speech Translator Project: Developed low- complexity online algorithms for embedded processors. Developed an online noise-robustness feature extraction algorithm, a binaural source-separation algorithm, an online VAD, an online MVN, and supporting technologies. Applied the algorithm to a 64,000-word database with a 500-hour training set.
• Voice Activity Detector project: Designed a robust VAD which requires very low computation

Graduate Research Work

• Power Normalized Central Coefficients (PNCC): Motivated by human auditory processing, this new feature set incorporates modulation frequency, temporal masking, and rate-nonlinearity concepts. The features require low computation and on-line implementation is possible.
• Small Power Boosting Algorithm (SPB): Developed the SPB algorithm motivated by the human rate-intensity as well as temporal and spectral smoothing. This algorithm works especially well for non-stationary noise such as music noise.
• Dual microphone speech enhancement systems: Developed the Phase- Difference Channel Weighting (PDCW) algorithm which performs sound source separation without a priori knowledge of the source locations.
• Single microphone speech enhancement systems: Developed the Power- function-based Power Distribution Normalization (PPDN) algorithm which enhances speech.
• Automatic Speech Recognition with low computational complexity: focused on developing online algorithms, which are robust against noise with low computational complexity.
• SNR Estimation algorithm for speech: Developed a new algorithm which estimates signal-to-noise ratio.
• Automatic ITD threshold detection algorithm: Developed a new algorithm which can obtain the optimal ITD treshold for sound source separation.
• Dereverberation algorithm Developed Suppression of Slowly-varying and Falling edges (SSF): a simple algorithm for robust speech recognition that is highly effective in reverberant environments.
• HLab C++ Automatic Speech Recognition System: Developed a C++ HMM (Hidden Markov Model) speech recognition system from scratch.
• Power Normalized Voice Activity Detection (PN-VAD): Developed a new voice activity detection algorithm that is very accurate and computationally efficient.