Evaluating MFCC Dimensionality and Machine Learning Classifiers for Speaker-Dependent Voice Authentication with Limited Training Data
Year: 2026
Type: conference-paper
Abstract: Voice biometrics have been a significant area of research due to their potential to provide improved levels of security when compared to traditional methods. In this paper, a comparative evaluation of different Mel-Frequency Cepstral Coefficient (MFCC) dimensionality ranging from 13 to 128, pairing with Support Vector Machine (SVM), Random Forest, and K-Nearest Neighbor models, was tested and compared as a speaker-dependent voice authentication system under low-data conditions. A total of 165 speech recordings were collected, and 30 used for internal test and 15 for external test evaluation. The experimental results show that all classifier-feature pairs performed perfectly on the internal test database. However, for the external test database, models performance range from 80.0% to 93.3%. The highest level of accuracy, at 93.3%, occurred when using SVM with 20 MFCC features. Overall result suggests that compact MFCC feature representations may be sufficient for speaker-dependent voice authentication under limited-data conditions.
Institution: University of Technology Malaysia
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Funder: Universiti Teknologi Malaysia
OpenAlex ID: W7225755161