Speech Intelligibility Measures for Speech Enhancement under Nonstationary Acoustic Noise

This paper presents the study of four objective intelligibility measures to evaluate the performance of speech enhancement techniques. The objective measures have high correlation with the speech intelligibility rates obtained in subjective listening tests. Four noise reduction algorithms are applied to enhance the speech signals corrupted by real highly nonstationary acoustic noises. The results show that, for most of the noise conditions, the best intelligibility results are obtained by the speech enhancement technique based on the empirical mode decomposition. Moreover, the Wiener filtering based technique also achieves interesting intelligibility gain for the highly nonstationary noises.