Abstract:A higher intelligibility subspace speech-enhancement algorithm based on the relative Root Mean Square (RMS) of speech segmental Signal-to-Noise Ratio (SNR) with low SNR is proposed. Under harsh conditions of low SNR, an improvement of noisy speech quality based on the majority existing speech-enhancement algorithms is often accompanied by a decrease in speech intelligibility. One important reason is that these algorithms only use Minimum Mean Square Error (MMSE) to constrain speech distortions but ignore that speech distortions caused by speech enhancement algorithms have different intelligibility influences on different speech segments. In order to overcome this disadvantage, the RMS of short-time segmental SNR was used to classify speech segments. Then the gain matrix components of middle-level RMS segments were modified to reduce the influence of speech distortion on enhanced speech intelligibility. Objective evaluation shows that the improved algorithm can improve enhanced speech intelligibility Normalized Covariance Metric (NCM) evaluation values. Subjective audition shows that the proposed algorithm does improve the enhanced speech intelligibility.