How do we perceive noisy and noise-reduced speech?

摘要

It is important for speech applications to quantify the intelligibility and the quality of speech. There are some means of evaluating the intelligibility and the quality objectively and subjectively. The objective evaluation is economical, but the general-use distortion measure is not available. On the other hand, the subjective evaluation requires a cost-consuming formal listening test. The right evaluation should be carried out in the right purpose. No almighty methodology has been established at the moment. In this paper, brain activities involved in evaluating the noisy and noise-reduced speech are measured using the functional magnetic resonance imaging (fMRI) technique. It is found that the brain activity in evaluating the noise-reduced speech varies depending on the noise reduction approach. The noise-reduced signals, which are obtained by the linear processing such as the delay-and-sum beam forming, calm down the brain activity compared with the noisy speech signals. On the other hands, non-linear signal processing such as the spectral subtraction activates the auditory cortex intensely. It is necessary for designing the advanced speech distortion measure to independently consider how the noise is removed and how the speech is distorted by noise reduction.