CRP Funded Projects

A.   Development of Algorithms and Systems for Robust Speech Recognition in Noisy Environments (Grant No: CRP.4.MAR11.4)

Summary - The prevalence of Automatic Speech Recognition (ASR) systems in the real world has been limited by their inability to perform adequately in the presence of noise. With an increase in the penetration of the ASR into many devices and control systems, the need for proper noise mitigation and reduction techniques have become a fundamental requirement to cope with the various kinds of noise present in the operating environment. Babble noise is one of the many classes of corrupting noise. This research seeks to extend the current base on two fronts. Firstly, the statistical distribution of babble as a function of the number of overlapping conversations as well as investigations into how babble interacts with the base units of clean speech, the phoneme, in the Mel Frequency Cepstral Coefficients (MFCC) feature domain and finally applying the knowledge gained from the first and second investigations to the existing method of spectral subtraction in the feature domain. Secondly, the proposal of a heuristic based approach for correction of babble corrupted noise. These investigations would improve the robustness of automatic speech recognition systems to the babble noise and help in improving the recognition accuracy of an ASR system. Currently we are using machine learning and artificial neural network techniques to further enhance the performance of automatic speech recognition system.

Output:

Dr. Jamin Atkins completed his Ph.D. degree in 2020.

We have published two papers in refereed international journals (10 and 12) and one national conference paper (7) out of this work. Two more papers are being submitted for publication in refereed journals.

Based on Jamin’s research work, he was hired by Digicel Group as Senior Regional Data Scientist (May 2015) and now working as Senior Advanced Analytics Specialist at National Commercial Bank Jamaica Limited.

Currently, Mr. Bryan Hibbert is continuing further research work in this area as an MPhil student.

B.   VLSI Implementation of various Digital Signal Processing Algorithms (Grant No: CRP.4.NOV08.1) and Development of Field Programmable Gate Array (FPGA) based Viterbi Decoder (Grant No: CRP.3.MAR09.21)

Summary: VLSI implementation using FPGA is a semicustom design approach using which one can quickly launch his application specific product in the fast growing market without expending millions of dollars in custom design process. In these projects following two highly demanding digital signal processing applications were implemented on FPGA:

Viterbi algorithm to design decoder for speech recognition applications.

FFT algorithm to design FFT Processor for spectrum analysis of data communication systems.

The ultimate aim of this work was to produce a very low cost hardware based implementation of a speech recognition system and spectrum analyzer, with an FPGA acting as a co-processor, which is capable of performing recognition and spectrum analysis at a much faster rate than software or DSP chip.