Automatic Identity Recognition Using Speech Biometric

Abstract

Biometric technology refers to the automatic identification of a person using physical or behavioral traits associated with him/her. This technology can be an excellent candidate for developing intelligent systems such as speaker identification, facial recognition, signature verification...etc. Biometric technology can be used to design and develop automatic identity recognition systems, which are highly demanded and can be used in banking systems, employee identification, immigration, e-commerce…etc. The first phase of this research emphasizes on the development of automatic identity recognizer using speech biometric technology based on Artificial Intelligence (AI) techniques provided in MATLAB. For our phase one, speech data is collected from 20 (10 male and 10 female) participants in order to develop the recognizer. The speech data include utterances recorded for the English language digits (0 to 9), where each participant recorded each digit 3 times, which resulted in a total of 600 utterances for all participants. For our phase two, speech data is collected from 100 (50 male and 50 female) participants in order to develop the recognizer. The speech data is divided into text-dependent and text-independent data, whereby each participant selected his/her full name and recorded it 30 times, which makes up the text-independent data. On the other hand, the text-dependent data is represented by a short Arabic language story that contains 16 sentences, whereby every sentence was recorded by every participant 5 times. As a result, this new corpus contains 3000 (30 utterances * 100 speakers) sound files that represent the text-independent data using their full names and 8000 (16 sentences * 5 utterances * 100 speakers) sound files that represent the text-dependent data using the short story. For the purpose of our phase one of developing the automatic identity recognizer using speech, the 600 utterances have undergone the feature extraction and feature classification phases. The speech-based automatic identity recognition system is based on the most dominating feature extraction technique, which is known as the Mel-Frequency Cepstral Coefficient (MFCC). For feature classification phase, the system is based on the Vector Quantization (VQ) algorithm. Based on our experimental results, the highest accuracy achieved is 76%. The experimental results have shown acceptable performance, but can be improved further in our phase two using larger speech data size and better performance classification techniques such as the Hidden Markov Model (HMM).