Speech presence probability estimation based on temporal cepstrum smoothing