High Frequency Speaker Identification Through the use of Automated Recognition Models

Jeremy Donai, Helen Boyd-Pratt, and Jillian Dodson*

Department of Communication Sciences and Disorders, West Virginia University, Morgantown, WV 26508

Presentation Category: Health Sciences (Poster Presentation #118)

Student’s Major: Communication Sciences and Disorders

The purpose of this speech classification study is to determine how efficient an automated recognition model, (i.e., a computer program), is at distinguishing person to person speech. More specifically the model will be using high-frequency energy at 4,000 Hz and above. Based on past research, it is hypothesized that the automated model should be able to distinguish speaker sex, identity, and understand the vowel sound that is being produced from this frequency range. The model will be given 7000 speech sounds produced by 25 male speakers and 25 female speakers. The important frequencies for identifying human speech, sounds below 4000 Hz, will be filtered out by a high-pass filter to allow for the computer to only receive a select high frequency range (at 4,000Hz and above). This experiment will use classification techniques such as kNN (k nearest neighbor) or SVM (support vector machine) in order to correctly classify the speakers. These programs are two types of supervised machines that are used to classify materials into separate groups. The finding of this study is expected to show a high correct percentage of speaker identification from the model. As well, it will show that the automated model can distinguish speech sounds without low frequencies. Implications of this study will allow for more advanced speaker identification, as well as important findings for the way humans and machines process speech.

Funding:

Program/mechanism supporting research/creative efforts: WVU's Research Apprenticeship Program (RAP) & accompanying HONR 297-level course