Client Recognition Using Text-Dependent Voice Biometrics: A Machine Learning and Deep Learning Approach

https://doi.org/10.5281/zenodo.21606822

Abstract

Human–machine interaction increasingly relies on natural spoken language, which makes robust and efficient speaker recognition a critical enabling technology. In this work we present a text-dependent speaker (client) recognition system built around the common Is-lamic greeting phrase “Assalam-Alaikum”, chosen so that the system is directly relevant to a large global user base. We collected a dataset of 500 speakers (250 male and 250 female), each contributing four utterances of the target phrase, for a total of 2000 raw audio recordings. The recordings were pre-processed to remove silence, normalized in du-ration, and augmented using shifting, stretching, noise injection, pitch shifting and Vocal Tract Length Perturbation (VTLP) to expand the dataset to 10,000 samples. Five acoustic feature families were extracted using Librosa: zero-crossing rate, root-mean-square en-ergy, Mel-spectrogram, Mel-Frequency Cepstral Coefficients (MFCCs, including first- and second-order derivatives), and Chroma STFT. The resulting feature set was encoded, scaled and used to train ten classical machine-learning classifiers (Decision Tree, Gaussian Naïve Bayes, Linear Discriminant Analysis, K-Nearest Neighbors, Support Vector Machines with linear/polynomial/RBF kernels, Logistic Regression, SoftMax Regression and XGBoost) as well as a four-layer Recurrent Neural Network (RNN) implemented in Keras. Logistic Re-gression and SoftMax Regression achieved the best classical accuracy at 96% on the held-out test set, while the RNN model achieved 97.33% test accuracy and 96.60% accuracy on a separate, previously unseen validation set. These results indicate that MFCC-based repre-sentations, combined with data augmentation and a modest deep-learning architecture, can deliver highly reliable text-dependent speaker identification even across a relatively large, class-balanced speaker population of 500 individuals.