Research On Speech Enhancement And Related Technologies In Low SNR Environments

Speech is the most direct and convenient way of communication but the clarity and intelligibility of speech signals are reduced because of the interference of noise which hinders the normal communication between people.The noise attaches serious disturbance on speech processing in low Signal-to-Noise Ratio(SNR)environments.Therefore,people need more advanced speech enhancement technology to denoise the noisy speech in order for the follow-up speech signal processing such as speech recognition.In summary,it is of great theoretical and practical significance to study the speech enhancement technology in the low SNR environments.The purpose of this paper is to suppress the interference of noise in speech in low SNR environments and to further enhance the effect of speech enhancement.Therefore,we start our study in speech endpoints detection and proposed an adaptive speech endpoint detection method based on Mel-Frequency Cepstral Coefficients(MFCC)distance.Then,aiming at the problem that the enhancement effect of the Wiener filter speech enhancement algorithm degrades due to the poor endpoint detection accuracy,the detection section of the Wiener filter speech enhancement algorithm is improved based on the accurate endpoint detection and finally the enhancement effect is enhanced.Then the speech enhancement using deep neural network(DNN)is studied to solve the problem of speech enhancement in more complex background noise environments.(1)Through the investigation of the development history of speech enhancement technology,we knew the development of speech enhancement technology and related speech processing techniques.Especially in addition to speech endpoints detection technology,we also did some research work.We introduced a variety of commonly used speech endpoints detection methods and also did some simulation for later comparisons.The main frameworks and implementation process of spectral subtraction,Wiener filtering and neural network speech enhancement are also described in detail.(2)In addition to the problem that many traditional speech endpoints detection methods can not guarantee the detection accuracy in low SNR environments,this paper proposes an endpoints detection method based on the MFCC cepstrum distance combined with spectral subtraction of multi-window estimation.In the speech detection phase,we first calculated the MFCC cepstral distance between each frame and the non-speech frame as a feature and the detection is done by a pair of thresholds.The pair of thresholds is decided by the the distance between the MFCC coefficient of the preamble non-speech frames and the average MFCC coefficients of the background noise.Therefore,this method can adaptively adjusted the thresholds in different noise environments to improve the speech endpoint detection accuracy.Experimental data shows that the endpoints detection method has higher endpoints detection rate in low SNR environments than other commonly used detection algorithms and it has good robustness to noise.(3)The speech enhancement quality of traditional Wiener filter speech enhancement algorithm is affected by inaccurate endpoints detection.To solve the problem mentioned above,an adaptive endpoint detection method based on MFCC cepstrum distance is used to estimate the noise sections more accurately of Wiener filter enhancement algorithm.Experiments show that the improved algorithm can improve the quality of speech.(4)In contrary to the problem that the above method is ineffective in non-stationary noise environment,a speech enhancement system based on neural network is set up and it makes the quality of the enhanced speech better.Next,the Wiener filtering speech enhancement algorithm is used to pre-enhance the training and test speech and then sent them to the neural network for training and testing.That is to say,the two speech enhancement methods are combined to further enhance the speech enhancement quality in complex noise environments with low SNR ratio.