A New and Efficient Feature Extraction Method for Robust Speech Recognition Based on Fractional Fourier Transform and Differential Evolution Optimizer

One of the main challenges in speech recognition is noise resistant feature extraction. In this paper, a new feature extraction algorithm, called Fractional and Adaptive Power Normalized Cepstral Coefficients Algorithm, has been proposed as a noise-resistant method for speech recognition. This proposed feature extraction method is based on a fractional short-term Fourier Transform. The selection of fractional conversion coefficient is important for proper analysis of multi-component signals like speech. Therefore, the proposed method obtains the optimum parameter of α for fractional Fourier Transform based on the noise class in the environment, adaptively by the Differential Evolution meta-heuristic algorithm. Moreover, TI Digit and Noisex-92 are used for evaluation of the resistance and accuracy of the recognition of the automatic speech recognition system. Simulation results show more resistance and higher recognition accuracy of the proposed feature extraction method rather than other methods in noisy and without noise environments. In the proposed ASR system, the Support Vector Machine (SVM) classifier with a nonlinear kernel has been used. Also, all the simulations are performed in MATLAB.