A Comparative Study Of Speech Enhancement Algorithm And Its Application In Feature Extraction

Language is the product of the development of human social civilization.In modern society,the role of speech is no longer limited to daily communication,but is applied in many aspects.However,in the process of practical application,the influence of environmental noise is inevitable,which seriously interferes with the acquisition and recognition of speech signals,and also causes the performance of many speech processing systems to be dramatically reduced.Therefore,we need to conduct speech enhancement processing for speech signals polluted by noise.With the maturity of deep learning theory and the development of artificial intelligence,more and more speech enhancement algorithms based on deep learning have emerged,however,due to the high requirements of data sets and models,the amount of computation is relatively large,it is still necessary to optimize and improve the traditional algorithm to improve the anti-noise performance of the existing product system,and it has practical application value.In practical application,since the integrity of the algorithm for preserving the speech time-frequency information and the power consumption in the hardware circuit should be considered,it is of great significance to compare and study various algorithms from the aspects of noise reduction effect and computation.As all kinds of speech enhancement algorithms have their advantages and disadvantages,and have different denoising capabilities for different noises,therefore,it is necessary to select appropriate speech enhancement algorithms according to different usage environments in practical applications.In this paper,a variety of speech enhancement algorithms are analyzed theoretically,including fundamental spectral subtraction algorithm and multiwindow spectral subtraction algorithm,modified spectral subtraction algorithm,adaptive filtering algorithm,wiener filtering algorithm,and modified wavelet packet speech enhancement algorithm.The fundamental spectral subtraction algorithm separates speech signal and noise from the frequency domain of speech signal,its principle is simple and the computation is small,it is an easy-to-implement speech enhancement algorithm.The multiwindow spectral subtraction algorithm is an improvement on the basic spectral subtraction algorithm.In the process of calculating the amplitude spectral value and the phase spectral value,a number of mutually orthogonal data windows are used to reduce the variance error.The improvement of the modified spectral subtraction algorithm to the fundamental spectral subtraction algorithm focuses on the power spectrum estimation of noise.the endpoint detection and Bark subband division which conforms to the human ear's auditory perception characteristics are introduced,making the algorithm more accurate for the power spectrum estimation of noise.The adaptive filtering algorithm uses the previous filter parameters to automatically adjust the current parameters to adapt to the unknown statistical characteristics of the signal and noise to achieve optimal filtering,and its performance far exceeds the fixed parameter filter.Wiener filtering algorithm is based on the optimal mean square error criterion to obtain the enhanced speech signal,and a linear time-invariant system is designed to make the output signal as close as possible to the desired signal.The modified wavelet packet algorithm is based on the principle of wavelet packet speech enhancement,and it improves the determination of wavelet denoising threshold,combining with the principle of human auditory perception,it tracks the distribution of wavelet packet decomposition coefficient of noisy speech signals more accurately.Then the corresponding research on the application of speech enhancement in feature extraction is done,taking the pitch extraction as an example,the harmonic summation algorithm for pitch extraction from the perspective of harmonics is introduced in principle and the algorithm is reproduced.The harmonic summation algorithm has good pitch extraction performance.In the experiment,different types and different intensity of noise,including stationary noise and non-stationary noise,were added to the pure speech signal.In the Matlab environment,various algorithms were used for speech enhancement,and the speech enhancement performance of various algorithms was compared and analyzed.The evaluation method of speech enhancement performance includes noise reduction effect and computation,among which the noise reduction effect includes objective and subjective index:the objective indexs are SNR gain and segmented SNR,and the subjective index is the accuracy rate of speech signal after speech enhancement in human hearing.For the statistics of computation,the same piece of voice is selected and the time of speech enhancement processing by the algorithm in the running stage of the code is counted,so as to explore the speech enhancement performance and the computation of different algorithms in different noise environments.Through experiment contrast study found that:first,in terms of noise reduction effect,each situation of noise environment,the adaptive filtering algorithm,modified spectral subtraction algorithm and modified wavelet packet algorithm,compared with the other three kinds of algorithm performance is better,there is little difference between the fundamental spectral subtraction algorithm and the multiwindow spectral subtraction algorithm,the wiener filter algorithm is the worst performance.Second,in terms of computation,since the fundamental spectral subtraction algorithm has the minimum computation,as a benchmark to quantify the other five kinds of algorithm,the operation time of the multiwindow spectral subtraction algorithm is the longest,which is 18.65 times that of the fundamental spectral subtraction algorithm,the second is the adaptive filtering algorithm,which is 11.43 times;the third is the modified wavelet packet algorithm,which is 9.4 times;the next is the modified spectral subtraction algorithm,which is 8.32 times;and the last is the wiener filtering algorithm,which is 6.77 times.Thirdly,combining with the comparison results of noise reduction effect and computation,the computation of the modified spectral subtraction algorithm is reduced by 1/3 compared with that of the adaptive filtering algorithm,while the noise reduction effect is roughly the same.In generally,the compromise between the noise reduction effect and computation is realized,and the power consumption is reduced while the noise reduction effect is guaranteed.Finally,the modified spectral subtraction algorithm is used to extract speech features,that is,the noisy speech signal with or without noise reduction is extracted by the harmonic summation algorithm for pitch frequency,the results were compared with the pitch extraction results of pure speech to test the improvement of speech enhancement in feature extraction.The experimental results showed that the modified spectral subtraction algorithm,as a pre-processing before pitch extraction,could well retain the accuracy of speech features.The above research results are valuable for the performance improvement of the practical speech signal processing system,but it still needs to be further tested and improved in combination with the hardware and software environment of the specific system.