Development of a biometric authentication platform using voice recognition

Abstract

Nowadays, several areas have turned to speech recognition, as they have witnessed its efficiency in maintaining the privacy of users and facilitating access and use of numerous applications, systems even establishments. Our work revolves around recognizing a form of speech from a single word, where the main objective is to identify as accurately as possible a set of predefined words from brief audio clips. a Speech Commands dataset consisting of 65 000 one-second-long statements of 30 short term, is applied. We use a Convolutional Neural Network (CNN) to classify representatives with two-dimensional convolutions on the audio waveform. Unlike many classic techniques with critical feature engineering, we benefit from the power of deep learning to comprehend the feature representation while training. The prototype achieves an acceptable accuracy rate on the validation set. Our model has provided satisfactory results during training with high accuracy compared to its pre-descendants. Still, room for errors exists in words exceeding the range of the training data and especially noisy samples.