Q-Gaussian based spectral subtraction for robust speech recognition