Speaker identification using angle amplitude transform

Speaker identification is the process of determining the person who performed the speech from a given audio recording, and it is widely used in various fields such as security systems. In this study, a comparative analysis of feature extraction methods used in the speaker recognition process was conducted. The commonly used methods in the literature, MFCC (Mel-Frequency Cepstral Coefficients) and LPCC (Linear Predictive Cepstral Coefficients), were examined, and in addition, a newly proposed method, Angle-Amplitude Transform (AAT), was applied for the first time. AAT computes angle and amplitude values from the peaks and troughs of speech signals and generates two-dimensional images. Features were extracted from these images using the Histogram of Oriented Gradients (HOG) method, and classification was performed with the K-Nearest Neighbors (KNN) and Support Vector Machines (SVM) algorithms. Furthermore, the images were directly classified using a Convolutional Neural Network (CNN). The LibriSpeech and VoxCeleb1 datasets were used in the study, and the data was divided into training and testing sets. Experimental results show that with first and second order AAT combined with CNN, accuracy rates of 93.2% and 91% were achieved on the LibriSpeech dataset, while 75.5% and 71% were obtained on the VoxCeleb1 dataset, respectively. Moreover, the hybrid CNN approach combining MFCC and first order AAT achieved accuracy rates of 97.6% on LibriSpeech and 84.3% on VoxCeleb1. The findings demonstrate that the AAT method provides an alternative and effective approach in the field of speaker recognition. Keywords: Speaker identification, feature extraction, angle amplitude transform, classification algorithms