Sistem Pengenalan Pembicara Menggunakan HMM Dan FFT

Abstract

This study analyzes the performance of a speaker recognition system based on Hidden Markov Models (HMM) using Fast Fourier Transform (FFT) for feature extraction. The objective of this research is to evaluate the effect of the number of FFT samples on the recognition accuracy. Experiments were conducted using speech signal test data with variations in the number of speakers ranging from 2 to 10, and variations in the number of FFT samples of 2, 4, 8, 16, 32, 64, 128, and 256 samples per state.The results indicate that the number of FFT samples significantly affects system accuracy. Using a small number of FFT samples results in low and unstable accuracy, especially as the number of speakers increases. Increasing the number of FFT samples improves both accuracy and stability of the system. The highest and most consistent performance is achieved when using 64 to 128 samples per state. Increasing the number of samples beyond this range does not lead to a significant improvement in accuracy, indicating a saturation effect. These findings demonstrate that selecting an appropriate number of FFT samples is essential for achieving optimal performance in HMM-based speaker recognition systems, with 64 to 128 samples per state recommended as the optimal configuration.