Deep Learning-Based Automatic Speaker Recognition Using Self-Organized Feature Mapping

Automatic speaker recognition (ASR) plays the major role in many applications including forensics, dictionary learning, voice verification, biometric systems, and so on. The performance of these application depends on efficiency of ASR system. However, the conventional ASR systems were developed using standard machine learning algorithms, which resulted in low recognition performance. Therefore, this work is focused on development of deep learning-based ASR system. Initially, voice features are extracted using Mel-frequency cepstral coefficients (MFCC), which analyzed the spectral properties of various voice samples. Then, self-organized feature map (SOFM) is applied to reduce the number of available features, which selects the best features using Euclidian similarity between features. Further, deep learning convolutional neural network (DLCNN) model is used to train the features and forms the feature database. Finally, a test voice sample is applied to the trained DLCNN model, which recognizes the speaker detail. The simulations carried out on Anaconda (TensorFlow) showed that the proposed ASR-Net system resulted in superior recognition performance as compared to conventional systems.