Objective Measures For Predicting Speech Intelligibility Based On Cochlea-saceled Entropy

The speech quality and speech intelligibility is a branch of the field of speech enhancement. The research on speech quality evaluation algorithm pays attention to the inhibitory process background of noise and other interferences, adopting the auditory method or understanding technique to improve the reliability of communication as well as the robustness of the system. The subject investigated by speech intelligibility is speech signal itself, differencing from speech quality, to enhancing speech signal from another perspective.Although the speech quality and speech intelligibility is a branch of speech enhancement, when it goes to the pronunciation listeners received,’speech quality’refers to the clarity of speech, in another word, the effect of other disturbance to speech, while’speech intelligibility’means the possible speech comprehension of audiences. The level of speech quality is not always match that of speech intelligibility. For instance, even though under high background noise, the listener can sometimes catch the meaning of speech expressions, while sometimes can’t in a quiet place. Therefore, this research, from the perspective of speech intelligibility, applies the method of extracting speech cochlea-scaled entropy to fragment voices and extract speech cochlea-scaled entropy, and then, uses normalized covariance measure and speech-to-reverberation modulation energy ratio algorithm to calculate objective evaluation scores. By four kinds of correlation degree analysis methods evaluating the correlation of subjective score and objective score respectively, objectively evaluates speech intelligibility.The main contents of this study are as follows:1Fragment voices, and extract speech cochlea-scaled entropy. Divide cochlea-scaled entropy into high entropy and low entropy.2Normalized covariance measure and speech-to-reverberation modulation energy ratio algorithm are adopted, high entropy speech and low entropy speech are put in. The distinguish of two algorithms are the latter one has got rid of the dependence of pure speech, so it is more suitable for daily use, such as artificial cochlear implantation and the wearing of hearing aids.The results show that speech intelligibility is higher when adopting high entropy speech after extracting speech cochlea-scaled entropy, illustrating that high entropy speech carries more speech information. In addition, the high entropy speech intelligibility correlation coefficients of subjective score and objective score, adopting normalized covariance measure and speech-to-reverberation modulation energy ratio algorithm, tend to oscillate at about0.85. However, normalized covariance measure is idealized for the impossibility of pure speech in real life. Speech-to-reverberation modulation energy ratio algorithm is more practical for needing no input of pure language.