Performance Evaluation of Speaker-Independent Automatic Speech Recognition System

摘要

Performance of automatic speech recognition systems severely degrades due to various sources of variability in speech waveform from various speakers. We attempt here to exploit the effect of different features on the automatic speaker independent speech recognition system accuracy. An Arabic isolated-digits recognition system based on Neural Network classifier, has been trained using a set of different features which include Fast Fourier Transform FFT. Spectrogram, Linear Predictive Coding LPC, and PLP. The results showed that the features which completely represent the speech signal ie. carry both the linguistic message and speaker dependent information, lead to decrease the recognition accuracy.