Date of Award
12-2024
Culminating Project Type
Thesis
Styleguide
apa
Degree Name
English: Teaching English as a Second Language: M.A.
Department
English
College
College of Liberal Arts
First Advisor
Ettien Koffi
Second Advisor
Edward Sadrai
Third Advisor
Douglas Gilbertson
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.
Keywords and Subject Headings
Speaker Verification, Voice Biometrics, Acoustic Phonetics, Native Hebrew Speakers, Vowel Analysis
Abstract
This experimental analysis investigates the robustness of 11 acoustic phonetic features of six vowels among native Hebrew speakers of English for speaker verification purposes, utilizing the Praat software for extraction. The research aims to determine whether these features are robust enough for effective speaker verification and to identify if any single feature can independently serve this purpose. Additionally, it explores which vowel is most discriminatory between the imposters and the person of interest (POI). This study measures the acoustic phonetic features from running speech, a method that is commonly used in speaker verification and impersonation studies. The speech samples for this analysis come from the Speech Accent Archive by Weinberger (2015). The participants are 11 native Hebrew speakers of English: seven males and four females. The English corner vowels /i, æ, u, ɑ / and the mid vowels /o, e/ are analyzed. The methodology used in the study consisted of extracting, measuring, and annotating the features from three lexical items in each vowel set. The extracted features are F0, F1, F2, F3, F4, Harmonic-to-Noise Ratio (HNR), jitter, shimmer, mean Mel-Frequency Cepstral Coefficients (MFCC), intensity, and duration. The study reveals that most of the examined features are sufficiently robust for speaker verification, with the exception of intensity, which is the weakest feature. Additionally, no single feature proves adequate for verification on its own; a combination of features provides the most comprehensive and reliable speech profile. The lot vowel, [ɑ], is identified as the most robust in the formant domain, with the trap vowel, [æ], also proving to be discriminatory. The spectral domain resulted in [i] being robust for HNR, jitter, shimmer, and [u] being robust for MFCC. These findings highlight the importance of using a diverse set of features to enhance speaker verification accuracy. An unexpected result emerged from the mean MFCC analysis, which did not align with findings from other features. This discrepancy suggests that while widely used in speaker verification, it should be evaluated alongside other features to ensure an accurate analysis. The [o] vowel was the most robust for the weakest correlate, intensity, and the [e] vowel was the most discriminatory for duration. One limitation of the study is the reliance on a single speech sample per participant. This in-depth and comprehensive research highlights the value of feature extraction via Praat in scenarios where automated speaker verification (ASV) systems are unavailable or unreliable. The findings contribute to forensic linguistics by providing detailed insights into acoustic features for multilingual populations and suggesting improvements for both Praat and ASV methods. Future research should incorporate more samples per speaker and different contextual factors. Integrating ASV technology could enhance the reliability and accuracy of speaker verification methodologies. A long-term study could also explore how verification features evolve over time in speakers of more than one language.
Recommended Citation
Rodgers, Megan C., "Experimental Analysis: Vowel Features of Native Hebrew Speakers of English for Speaker Verification" (2024). Culminating Projects in TESL. 71.
https://repository.stcloudstate.edu/tesl_etds/71
