Research On Low SNR Speech Detection And Enhancement Methods Based On HHT In Complicated Environment

With the development of information processing technical means and the extension of speech signals application field, speech signals processing have been increasingly required. Speech recognition technology requires to achieve effective recognition in the background noise. Speech communication technology requires to decrease data transmission in the no speech band as much as possible without affecting speech signals quality received. Intercepting opponent information and transmiting information effectively are the important component in the military. However, in the complicated acoustics environment, the speech signals of above situation became weaker relatively due to the effect of strong background noise, such as speech communication and recognition in the noisy environment outside, information interception in the HF communication with the stronger noise etc.The situations have usually appeared. Therefore, it has significant meaning to detect and enhance speech signal with low SNR effectively in the complicated background environment.Time-frequency analysis is one of the effective means in the speech signals detection and enhancement.Traditional time-frequency analysis algorithm have short-time fourier and wavelet transform etc. The algorithms need to set signal decomposition scale artificially by experience.However, the scales doesn't sometimes response signal characteristics. Furthermore, wavelet transform need to choose wavelet base beforehand by transform object, whereas fixed wavelet base has not fit the whole signal analysis. According to this, the article selected basic function and filter scale algorithm adaptively by signal characteristics, Hilbert-Huang Transform. The algorithm is specially proper to analyze non-stationary signals.We researched speech signal detection and enhancement with low SNR based on HHT. The main achievements includes these contents.Firstly, the paper adopted a speech signal detection algorithm with low SNR based on HHT for solving the speech signals detection with low SNR in Gauss environment. The core of the algorithm is to separate IMFs of more noise in the speech signals with low SNR by adaptive decomposition characteristic of EMD. According to other IMFs constructing Hilbert energy spectrum, speech signals with low SNR have been detected.Secondly, the paper presented a detection algorithm of speech signals with low SNR based on HHTSM for solving the speech signals detection in many environments. The core of the algorithm is to realize speech signals detection by the difference of HHT spectrum between speech signals and noise signals. Constructing HHTSM of speech signals with low SNR, energy concentration has been considered, so the paper adopted a frame as unit and constructed spectrum matrix of the signal. The article presented a 3D viewable analysis method for analyzing spectrum matrix effectively. It effectively found the difference between speech and noise in the spectrum matrix, and implemented weight filter to the matrix by the difference setting filter coefficient. Spectrum matrix has been transformed 2D time-amplitude curve frame as unit after filter, and it calculated threshold adaptively. Speech band has been detected.Then, the paper adopted EMMD algorithm to improve curve fitting and endpoint effect of EMD. The core of the algorithm is to construct mean curve by the whole information of signals. Constructing mean curve of signals, integral mean value theorem has been applied by all datas among the extreme points. Practical value has been found in the a signal as means. Obtained means responsed the true means of input signals. It improved estimation accuracy of local mean, and reduced endpoint effect.Additionally, the paper presented a speech signal with low SNR algorithm based on EMMD for solving the speech signals enhancement with low SNR in any environments. The core of the algorithm is to filter after screening decomposited IMFs, preventing over-filter and owe-filter. The article presented a maximum similarity judgement algorithm by the maximum similarity of noise signals and IMFs. As noise and speech signals easily spectrum aliasing, we adopted time domain filter by the maximum similarity screening IMFs. The new signal has been reconstructed by IMFs which has been filter and IMFs which has been no processing. Then enhancement results have been obtained.Eventually, it presented a speech signal enhancement algorithm with low SNR based on EMMD and ICA for solving the speech signals enhancement with low SNR in more environments. The core of the algorithm is to separate intrinsic properties of IMFs and to eliminate information confusion by ICA. To improve maximum similarity, IMFs which have been screened are processed by ICA. It was beneficial to concentrate noise properties and eliminate noise. ICA after filter must be twice reconstruction, which are ICA reconstruction and EMMD reconstruction, as the unconcerned amplitude and location of ICA. The enhancement results have been obtained.The dissertation presented relevant algorithms aimed to different noise environments, with detection and enhancement of speech signals with low SNR by HHT characteristics of speech signals with low SNR in the complicated environment. It analyzed properties and experiments of the algorithms. Finally, the speech signal detection and enhancement with low SNR have been solved.