Noise suppression based on auditory-like filters for robust speech recognition

An efficient noise suppression algorithm for speech signals is presented; it is based on auditory-like filters. The algorithm processes in three steps. First, the noise spectral energy is estimated after corrupted speech is input into a set of auditory-like filters. A statistical estimation method based on multi-band filters is proposed and compared with the weighted average. The second step is to eliminate the estimated noise spectrum from the observed signal by spectral subtraction. Finally, an auditory-based feature is extracted from the enhanced signal and introduced into the ASR system. The noise suppression algorithm is evaluated in a speaker-dependent Chinese digit experiment and experimental results show that the proposed algorithm can improve the automatic speech recognition performance in a noisy environment.