Speech Enhancement Method Improving Speech Intelligibility Effectively

The state-of-art speech enhancement methods are capable of suppressing thebackground noise and improving speech quality, but they are not able to improvespeech intelligibility in low Signal-to-Noise Ratio (SNR) conditions. In order to solvethis problem, based on the research of traditional statistical-model based methods, aspectral amplitude estimator based on improved β-order Weighted EuclideanDistortion Measure (I-β-WEDM) is proposed in this paper, which could improvespeech quality in high SNR conditons. Furthermore, based on the effect of traditionalspeech enhancement methods on speech intelligibility, a speech enhancement methodusing the Constraints on Speech Distortion and Noise Over-estimation (CSDNO) isproposed, which could improve speech intelligibility effectively in low SNRconditions.The main research work is embodied in the following aspects:First, the statistical-model based speech enhancement algorithms are studied. Theβ-order Minimum Mean Square Error (β-MMSE) estimator could control the amountof noise reduction by choosing the order β adaptively, but the residual noise in speechperiod is too much. On the other hand, by applying different emphasis on the errorsnear spectral peaks and valleys, the WEDM estimator could suppress the noise moreeffectively in spectral valleys, but the amount of noise reduction can not be controlledadaptivly. Combing the adavantages of the two estimators, a novel speechenhancement algorithm based on I-β-WEDM is proposed, in which, a cost functionwith the form of β-order WEDM is introduced and the order is updated according tothe SNR in each critical band. The evaluation of the proposed method is performedunder ITU-T (International Telecommunication Union, TelecommunicationStandardization Sector) G.160. The results indicate that, comparing with the referencemethods, the proposed method can produce larger amount of noise level reductionwith better objective speech quality.In addition, we reviewed why the existing speech enhancement method failed toimprove speech intelligibility, and the effects of speech distortion and noiseover-estimation on the speech intelligibility are analyzed. According to the researchand the IdBM (Ideal Binary Mask) criteria, a speech enhancement algorithm forspeech intelligibility improvement is proposed through the modification of noiseestimation and gain function based on CSDNO. In order to further suppress themusical noise, based on the different impacts of various speech distortions on speechintelligibility, an improved CSDNO (ICSDNO) method is proposed through themodification of SNR estimation. The proposed methods could effectively improve thespeech intelligibility in low SNR conditions. Finally, subjective and objective tests are performed on the proposed methods.The performance evaluation includes two aspects: G.160test and the speechintellibility test. The results of G.160test indicate that, comparing with the referencemethods, the proposed methods have less distortion on speech level, and the speechquality satisfy the requirement. The speech intelligibility test includes the objectivetest based on frequency-weighted SNR segmental (fwSNRseg) and fractionalArticulation Index (fAI), and the subjective test based on Determine Rhyme Test(DRT). The results show that, comparing with the reference methods, the proposedmethods could improve speech intelligibility effectively with less speech distortion.Additionally, the proposed methods are realized using fixed-point C language.