Speech Frequency Parameter Estimation In Low SNR

Low SNR voice frequency parameter estimation has been a very important research direction of the speech signal processing.Pitch parameters and formant parameters are very important theoretical and practical significance for voice recognition, speech synthesis and voice compression coding.This paper studies from the basic theory of the study of speech signal processing to the voice signal generation model with linear prediction analysis, anti-noise theory and method, linear prediction estimates in the low signal-to-noise ratio to improve to improve the extraction of speech fundamental frequency and formant frequency low signal-to-noise ratio, and made certain improvements in the pre-and post-processing of the speech signal. The results show that the new method can get a smaller linear prediction error and improve the accuracy of the linear prediction by Matlab platform.The main research work are as follows:First,this article analyzes the generation and characteristics of the speech signal, speech signal generated model and the theory of linear predictive analysis, pre-emphasis and framing a common method of pre-and its endpoint detection analysis. Considered to improve the strike accuracy of the linear prediction coefficients, the linear prediction analysis method using covariance lattice method, and a more detailed analysis principles. In addition, the probability distribution of the voiceless and voiced signal on the zero-crossing rate differences, combined with the good performance of the autocorrelation function assay voiced signal, propose a suitable low-the voiced signal-to-noise ratio endpoint detection method in Matlab platform the experiment, the better.Second, the common method of analysis of the pitch frequency and the formant frequency of the speech signal detected were discussed. Fundamental frequency detection, analysis and comparison of the auto-correlation function, the average magnitude difference function method, the cepstrum and simple inverse filtering method. Low signal-to-noise ratio, the auto-correlation function prone octave or half-frequency phenomenon, less effective detection, cepstrum and the average magnitude difference function serious decline in performance, simple inverse filtering method is better than the performance of the other three methods. Common method for formant frequency detector, the short time Fourier transform method is affected by noise, low signal-to-noise ratio for formant estimation error; cepstrum can realize the separation of channel and incentive, but generally cepstrum envelope points less accuracy is not high, and the cepstrum large amount of computation, seriously affected by noise, low signal-to-noise ratio is not suitable for testing; linear prediction method compared to the other two method is effective, more practical.Then, this paper analyzed and discussed the impact of noise on the signal, as well as the principles and methods of noise reduction. When taking into account the order of selection of the spectral envelope of the linear prediction, when the number of lower order linear prediction the spectral envelope more slip is inaccurate, higher order linear prediction (LPC), and the resulting spectral envelope subject baseband serious and often relatively sharp, and based on the amplitude spectral envelope of the speech signal to strike the linear prediction coefficients can be well overcome the error of prediction coefficients Select Select brought. Therefore, this article discusses the use of the amplitude spectrum of the speech signal after the noise reduction to strike the linear prediction coefficients, and optional co-variance lattice method for LPC analysis to improve the accuracy of the prediction coefficients, and using the obtained LPC coefficients to inverse filtering strike pitch frequency, spectral envelope analysis of the to strike first three formant frequency. And voice frequency parameters obtained by this method and the simple inverse filtering method to strike a pitch frequency and linear prediction method to strike a formant frequency compared to the Matlab simulation shows the false detection rate in the low signal-to-noise ratio, better performance. Further, the post-processing of the speech signal, the Q analysis and discussion of a median filter method, taking into account the voice frequency detection may occur in the wild point, with a median filtering method discussed in after strike a speech frequency parameter, filter can be obtained relatively smooth voice frequency parameter trajectory.Then, we designed a software, through Matlab platform GUI design interface, that can detect the pitch frequency and the first three formant frequencies. The software can be two ways to upload voice files and live recordings from the local voice data, test results are displayed in the interface by selecting the appropriate control options, and can choose a frame of data to specific parameter data is displayed in the right side of the interface.Finally, there is a summarize of the full text work done. And introduced voice frequency parameter detection prospects for future research directions.