A Modified A Priori SNR Estimator Based On The Speech Reassigned Spectrogram

In daily speech communication, speech signal inevitably interfered by different kinds of noise, which seriously reduce the speech intelligibility and clarity, so the study of speech enhancement technology to reduce noise is very necessary. The purpose of the speech enhancement is as much as possible to restore the speech signal contained noise into pure signal, so as to improve the quality and intelligibility of speech. Speech enhancement algorithm is widely used in video speech communications and digital hearing aid, etc.The traditional speech enhancement algorithms can be divided into single channel and multi-channel speech enhancement algorithm according to the principles of speech enhancement algorithm. The calculation of the single channel speech enhancement algorithm is small, so it is easy to implement. The existing priori SNR algorithm can remove noise effectively, enhancing the performance of speech enhancement algorithm, but it only considers the characteristic between speech frame while ignores the language features between audio, which lead it unable to use speech frequency correlation to restrain non-phonetic component. According to this problem, this paper presents an improved estimate of the a priori SNR, which is based on the Reassigned Spectrogram(RS: Reassigned Spectrogram Algorithm). In the RS algorithm, we take advantage of the Channel Instantaneous Frequency(CIF: Channel Instantaneous Frequency) and Local Group Delay(LGD: Local Group Delay), meanwhile, it consider the features of speech signals between frame and frequency. Simulation results show that both can improve estimator algorithm performance on priori SNR, but also found that the RS algorithm has a problem of speechdistortion with low SNR Babble noise. So, in this paper, we use the cepstrum processing to restrain harmonic frequency component in speech, accurately estimate the noise power spectrum,reducing speech distortion and residual noise. In this paper, the main content and innovation points include the following three aspects.1) According to the classic priori SNR and speech signal time-frequency distribution, we go on the research and simulation analysis and summary. We found that the current priori SNR estimator algorithm has some defects, so this paper provide the basis of Reassigned Spectrogram method.2) According to study and analysis the characteristic of CIF and LGD, we use CIF to update the speech signal on priori SNR of current frame and current frequency band and we use LGD to update priori SNR estimator on the current frame on different frequency bands,putting forward a priori SNR estimator algorithm based on Reassigned Spectrogram.3) According to the test of the objective evaluation index, it indicate the RS algorithm proposed in this paper has some deficiency to nonstationary noise, through the cepstrum processing to let the RS accurate estimate the noise power and avoid the problem of noise power over estimator or underestimate, thus it can reduce speech distortion and residual noise to improve the performance of the algorithm. The experimental results show that the RS algorithm after cepstrum processing(RS_CP), has a low speech distortion and retain the less residual noise at the same time.This paper has the following two points on innovation:1) Combined use of CIF with LGD which embody the speech between frame and frequency characteristics, better able to reduce the residual noise.2) Using the cepstrum processing can better inhibit speech harmonic frequency components, to avoid the noise power spectrum over estimate or underestimate.