Research On Algorithms Of Speech Enhancement Based On Fractional Fourier Transform

Speech signal is one of the most important intercommunication manners among human- kind. But in the process of communication, speech signal is always disturbed by various kinds of noises from environments, which may lead to unreasonable results. Speech enhancement emerges as the times require. It deals with the problems of how to improve the signal quality, articulation and intelligibility and develop the performance of communication system.The traditional enhancement methods are mainly based on the properties of speech signal in time or frequency domain. In this thesis, speech enhancement technology is re- searched based on fractional Fourier transform, which contains the model of fractional spectral subtraction, the optimal transform order, speech endpoint detection and noise estima- tion.Firstly, the linear optimal operator is researched in fractional Fourier domain, and an improved speech enhancement algorithm based on classical spectral subtraction and linear optimal operator is proposed. Here we take the fractional Fourier transform on the premise of given orders, then extract the'given'amplitudes as the objects of noise estimation. The noise estimation is researched based on parameter adaptive adjustment. The simulations show that an optimal order exists in the interval of transform.Secondly, the spectrums of speech and noise, which are modeled as Gauss distribution is proposed. So we deal with the problem with the rule of minimum mean square error. Then we may obtain the best transform order of fractional Fourier transform, and treat with the amplitudes by fractional noise estimation in order to get the analogical amplitudes of original speech signal.Thirdly, noise estimation is researched based on iterative processing in fractional Fourier domain. Since noise estimation is related to speech endpoint detection, so an improved speech endpoint detection algorithm is proposed in this paper. The improved method is re- searched based on the method of high resolution computation of fractional Fourier transform, which performances more effective due to simulation results. High resolution computation is a novel method proposed recently.Finally, speech enhancement based on neural network is proposed, whereas the noise estimation is not high robust. If the fractional amplitudes of noise are higher than the current frame's amplitudes, then some characteristics of speech signal may lose in the process of spectral subtraction. So speech amplitudes are extracted from fractional Fourier transform as the training input of neural network, in order to get the optimal amplitude approximations. The algorithm not only receives the optimal approximation, but also can overcome un- certainties in signal processing. The simulation results show that the method performances excellent for white noise, as well as colored noise.