Hopfield Neural Network with Lotus Effect Optimization for Automatic Speaker Recognition in Noisy Environment
摘要
Speech is the most effective way for people to communicate with each other. It may transmit rich and important information including accent, gender, emotion, and distinct identity. A key property of a good automatic speaker recognition (ASR) system is being able to model the uncertainty and variation of the utterances of the same speaker. Speaker recognition under noisy and unconstrained conditions is an extremely challenging. Machine learning is one of the most widely used systems for speaker identification. But achieving better accuracy is difficult due to the inaccurate extraction of features. To overcome these drawbacks, optimized Hopfield neural network (HNN) is developed. Raw audio signals are preprocessed using Notch filter and adaptive thresholding to enhance the quality of the original signal. Then, extracting the features from the preprocessed signal using bark frequency cepstral coefficients (BFCC) as well as generalized perceptual linear prediction (GPLP). In the process of feature extraction, preprocessed signal is decomposed into frequency subbands and further normalized to obtain features coefficients. Extracted features are given as input for the classification process. Here, Hopfield neural network is employed for identifying the speakers in noisy environment. Lotus effect optimization algorithm (LEOA) algorithm is utilized for tuning the hyper parameters in HNN which is used to enhance the classifier performance. Proposed model achieved 94