Research On Speech Enhancement Method Based On Dual Microphone System

With the rapid development of science and technology,speech enhancement technology has been widely used in various speech communication systems and artificial intelligence products.In the field of speech enhancement technology,methods based on deep learning have been successfully applied to speech enhancement and achieved remarkable results.However,such methods are often highly dependent on data sets,and generally have high Computational complexity,which is not applicable to some application scenarios with limited computing resources.Among the classic speech enhancement algorithms,compared with the single-channel speech enhancement technology,the speech enhancement technology based on dual microphones can extract and utilize the spatial characteristics of the target sound source,so it has higher design flexibility and better enhancement effect;Compared with the speech enhancement technology based on the microphone array,the speech enhancement technology based on dual microphones has the advantages of small scale,low cost,and easy integration.Therefore,the speech enhancement algorithm based on dual microphones is more suitable for portable devices such as mobile phones and hearing aids.However,compared with single-channel speech enhancement and multi-channel speech enhancement methods,the current research on dual-microphone-based speech enhancement methods is not rich.Therefore,this paper firstly investigates and analyzes the existing classical speech enhancement algorithms,and then focuses on the noise suppression and de-reverberation methods based on dual microphones.Based on the existing achievements and technologies,this paper carries out some research work in the following aspects.(1)In view of the speech enhancement requirements of mobile phones,the classical dual-mic noise reduction algorithm based on power difference is analyzed,and the limitations of this method are summarized.(2)Aiming at the limitation of the dual-mic noise reduction algorithm based on power difference,a dual-mic noise reduction method based on the difference of signalto-noise ratio is proposed.The algorithm first estimates the existence probability of pure speech according to the posterior signal-to-noise ratio difference between the two microphones;at the same time,a transient noise capture module is introduced to estimate the existence probability of transient noise;then according to the existence probability of speech and transient noise The noise power spectral density(PSD)in noisy speech is estimated,including transient noise PSD estimation and steady-state noise PSD estimation;finally,based on the obtained noise PSD estimation,the timefrequency gain is calculated according to spectral subtraction.The main channel noisy speech signal is enhanced.Compared with the dual-mic noise reduction algorithm based on power difference,the designed dual-mic noise reduction algorithm based on the difference of signal-to-noise ratio has the following advantages: First,the probability of speech existence is calculated according to the difference of the posterior signal-to-noise ratio of the received signals of the two microphones,reducing the sensitivity difference between the two microphones affects the error of speech presence state estimation.Second,the estimation scheme of noise power spectrum is improved,and a transient noise capture mechanism is introduced,so that the algorithm is still applicable under transient background noise.(3)Under different types of background noise,the performance of the dual-mic noise reduction algorithm based on power difference and the dual-mic noise reduction algorithm based on signal-to-noise ratio difference are compared through simulation experiments.The latter has better improvement effect and applicability to acoustic scenes where transient background noise exists.(4)In order to eliminate the interference of reverberation on speech quality,an adaptive de-reverberation algorithm based on Weighted Prediction Error(WPE)is analyzed and applied in dual-mic scenarios.(5)Aiming at the problem that the calculation complexity of the WPE-based adaptive de-reverberation algorithm is too high,a low-complexity de-reverberation algorithm based on WPE is proposed.All of them obey the assumption of zero-mean time-varying variance Gaussian distribution.The sparse block diagonal matrix is used to approximate the inverse correlation matrix of the reverberated speech signal,which effectively reduces the computational complexity of the matrix operation in the original algorithm.(6)Through simulation experiments,the performance of the WPE-based dereverberation algorithm,the WPE-based low-complexity de-reverberation algorithm and other de-reverberation algorithms are compared,and the WPE-based dereverberation algorithm and its low complexity are verified.The de-reverberation algorithm can more effectively suppress the reverberation components.And through the actual measurement of the RK3399 platform,the statistical comparison of the WPEbased adaptive de-reverberation algorithm and the real-time running rate of its lowcomplexity algorithm has verified the low-complexity WPE-based voice processing performance under the premise of ensuring the reverberation speech processing performance.The de-reverberation algorithm can effectively reduce the required computational overhead.