Abstract: This paper concerns techniques of speech processing using a small-size microphone array to improve automatic speech recognition(ASR)performance in an indoor reverberant environment.The method first applies Estimation of Signal Parameters via Rotational Invariance Techniques(ESPRIT)to compute directions-of-arrivals of wideband speech signals and implements time-delay compensation accordingly;array signal filtering is then considered jointly with the HMM-based speech recognition procedure,whose outputs are fed back to the front end to optimize array filtering design.Different from conventional array processing aiming to enhancing signal waveform,the approach here adjusts array filtering coefficients to reduce mismatch between the features to be recognized and the training model,thus directly maximizing the likelihood of the right transcription for a selected vocabulary.Experimental data processing shows that the above approach can effectively reduce ASR error rate for an isolated word vocabulary of finite size in a meeting-room environment,superior to conventional beamforming processing;compared to a local optimization scheme,applying global optimization in array filtering design further improves the performance.