Experiences

Undergraduate Thesis

Chittagong University of Engineering and Technology, Chattogram, BD

• Machine-Learning Based Speaker Identification System Using Cochleagram.
• Objectives
  • To build a speaker identification (SID) system that provides consistent performance for different noises at various levels of SNR.
  • To make a robust SID system for text-dependent as well as text-independent dataset.
  • To evaluate the variation of the system's performance for reverberated speech as well as distorted speech.
• Abstract
  • The process of recognizing someone based on one’s voice is called speaker identification (SID). Speech signals being susceptible to significant variations, SID is a quite challenging task and conventional speaker identification systems does not perform well under different noisy environments. Also, reverberant conditions, recording quality and transmission method aggravate the problem. In this study, a robust system for SID is proposed using an auditory-inspired feature called cochleagram. Cochleagrams were generated using gammatone filterbank having 128 channels from frequency 50 to 8000Hz. Cochleagrams constructed from clean and a particular noise added over speech samples at a certain SNR were used to train a convolutional neural network (CNN) referred to as noise adapted CNN. The proposed model was then tested for 9 other different noises at different level of SNRs to measure the accuracy in mismatched conditions. Experimental results showed that the identification accuracy obtained from the proposed system is akin to that of the existing speaker identification (SID) systems. However, the proposed system showed quite better performance than existing neurogram based methods under noisy conditions particularly at very low SNRs for text-dependent as well as text-independent corpus.

Project Mentor

University of Iowa, Iowa City, IA, USA

View

• Evaluation of Different Bias Field Correction Approaches for Pulmonary UTE MRI.
• Objectives
  • To eliminate the presence of intensity inhomogeneities in pulmonary MRI for better diagnosis, plan and treatment of patients with lung disease.
  • To help facilitate various post-processing steps such as - lung segmentation.