Discipline: Technology and Engineering
Subcategory: Computer Science & Information Systems
Abelson Abueg - East Carolina University
Co-Author(s): Nasseh Tabrizi, East Carolina University, Greenville, NC
The goal of speaker recognition is to be able to recognize and identify speakers by their voice. From that, speaker recognition can be broken down into two smaller subtopics or subproblems, speaker identification, and speaker verification. A speaker is identified by a speaker recognition system by using characteristics parameters extracted by the speaker’s speech signals. In this paper, IBM Auto AI was utilized for speaker recognition using MFCC and GFCC features extracted from the TIMIT dataset and a generated ‘Noisy TIMIT’ dataset to test for robustness between the two speaker features and identify the best machine learning algorithm determined by IBM Auto AI. This research is important to determine if there is a significant difference between GFCC and MFCC with robustness when used with traditional machine learning algorithms. Two Auto AI experiments were conducted and Auto AI selected decision tree classifier as an optimal machine learning algorithm for both Mel-frequency Cepstral Coefficient (MFCC) and Gammatone Frequency Cepstral Coefficient (GFCC) features. The results of the experiment have shown that decision tree classifiers perform quite well as a robust algorithm for speaker recognition for both features. There has been past research in utilizing decision trees for speaker recognition with MFCC features, but the results of this research suggest that decision tree classifiers could work with other speaker features.
Funder Acknowledgement(s): This research was primarily supported by NSF REU grant, #216990.
Faculty Advisor: M.H. Tabrizi, firstname.lastname@example.org
Role: In my REU, we were to do everything ourselves to get the full experience of what research is like. I did everything: literature review, data preparation, research problem, experiment design, implementation, and writing up my paper report.