Speech intelligibility improvement in noisy reverberant environments based on speech enhancement and inverse filtering

Abstract

The speech intelligibility of audio systems in enclosed space is degraded by reverberation and background noise. This dissertation proposes a preprocessing method that combines speech enhancement and inverse filtering to improve the speech intelligibility in such environments. The influence of reverberation on speech intelligibility was investigated firstly. Based on the fast inverse filtering technique and normalized-LMS algorithm, a new adaptive room equalizer which considered the frequency response of cochlear basilar membrane was proposed to achieve the better equalization performance than the classical methods. In the stage of theoretical research, a single position adaptive equalizer is designed first to verify the equalization performance through MATLAB simulation. After that, a multi-position equalizer is proposed to equalize a small area instead of a single point. In order to verify the effectiveness of the method, the experiments were performed in three different rooms. The objective evaluation results from the experimental data illustrated that the adaptive equalization method based on the auditory model could further improve the dereverberation performance in the different reverberation time conditions. Furthermore, the subjective listening test also confirmed the effectiveness of the proposed method. The influence of background noise on speech intelligibility was investigated in the follow-up study. A transient speech enhancement method was used to reduce the noise masking under the different SNR and noise type conditions. A real experiment was carried out in an anechoic chamber and proved the effectiveness of this method. Based on the previous research work, a new preprocessing method combining auditory-model-based inverse filtering and transient speech enhancement algorithms was proposed for improving the speech intelligibility in noisy reverberant environments. The combination method could not only equalize the speech distortion caused by sound reflections in enclosed rooms but also reduce the noise masking in noisy environments. In order to prove the effectiveness and stability of this combination method, the real experiments were carried out in various noisy, reverberant environments, and the test results verified the effectiveness of the proposed method in different noisy reverberant conditions. In addition, a listening test was carried out to compare the performance of different algorithms subjectively. The objective and subjective evaluation results reveal that the speech intelligibility is significantly improved by the proposed method.