A machine learning based speech enhancement method is proposed to improve the intelligibility of whispered speech. A binary mask estimated by a two-class support vector machine （SVM） classifier is used to synthesize the enhanced whisper. A novel noise robust feature called Gammatone feature cosine coefficients （GFCCs） extracted by an auditory periphery model is derived and used for the binary mask estimation. The intelligibility performance of the proposed method is evaluated and compared with the traditional speech enhancement methods. Objective and subjective evaluation results indicate that the proposed method can effectively improve the intelligibility of whispered speech which is contaminated by noise. Compared with the power subtract algorithm and the log-MMSE algorithm, both of which do not improve the intelligibility in lower signal-to-noise ratio （SNR） environments, the proposed method has good performance in improving the intelligibility of noisy whisper. Additionally, the intelligibility of the enhanced whispered speech using the proposed method also outperforms that of the corresponding unprocessed noisy whispered speech.
Some factors influencing the intelligibility of the enhanced whisper in the joint time-frequency domain are evaluated. Specifically, both the spectrum density and different regions of the enhanced spectrum are analyzed. Experimental results show that for a spectrum of some density, the joint time-frequency gain-modification based speech enhancement algorithm achieves significant improvement in intelligibility. Additionally, the spectrum region where the estimated spectrum is smaller than the clean spectrum, is the most important region contributing to intelligibility improvement for the enhanced whisper. The spectrum region where the estimated spectrum is larger than twice the size of the clean spectrum is detrimental to speech intelligibility perception within the whisper context.
