A SOVF Speech Prediction Model In Hidden Phase Space Based On Improved DPSO Algorithm

With the development of artificial intelligence,speech recognition and prediction have become one of the basic tasks of some applications,such as intelligent control,individual identification,and data analysis.It has drawn extensive attention in recent years which the Volterra model can analyze nonlinear system in both the time and frequency domains to solve speech recognition and prediction problems.It can build up models to analyze nonlinear systems,and uses both linear and nonlinear features.Volterra model as a series function expansion could describe the relationship between the input and output of the system.It can work as a mathematical model to approach the characteristics of the original chaotic system.Usually,in reality,we replace the kernel function with other proper function,after replace the kernel function,it will have a better description of some nonlinear system.The original Volterra model needs to set the parameters of the model in advance,and a large number of work are required to prediction,especially for some speech time series that are difficult to set the parameters.In original Volterra model need to set the embedding dimension and delay time of the sequence at least,we also have to import the parameter into the model.If the parameters are set incorrectly,Volterra model will get the wrong kernel function.In the original Volterra model,if the embedding dimension of the speech time series is too large for the model,the number of kernel functions will be increase explosively,the parameters which have little influence to the system will always take part in the calculation and waste the resource of CPU.More seriously,too much calculation will make the parameter to be updated slowly or even cause dimensional disasters.The original Volterra model cannot illustrate the structure of the model explicitly during the model construction process.In the meantime,it cannot figure out the relatively large kernel function coefficients.Although the search speed of traditional Volterra model is faster than that of the general model in the model constructing process,it has a weak global search ability and cannot reach a high search accuracy.In this thesis,in order to tackle these weaknesses mentioned above,we improve the traditional Volterra functions and proposed a non-linear prediction model of chaotic speech signal based on the hidden phase space reconstruction.(1)In order to overcome the problem that the traditional Volterra model has a low operating efficiency and requires a large amount of complex preprocess,we introduce a second-order Volterra model based on the implicit space reconstruction.Compared with the traditional method,In the proposed method avoids the phase of space reconstruction and the SOVF model solution process(hidden space reconstruction)algorithm in the speech time series is implicitly contained the phase of space reconstruction process.The optimal embedding dimension and delay time are obtained in the process of the algorithm to solve the model coefficient.By utilizing the proposed method avoids setting the basic parameters of the sample in the model in advance,which enhances the adaptability of the model and improves the operating efficiency.(2)Considering that the traditional Volterra algorithm has slow update speed,low global search ability and low search accuracy,it cannot jump out of the local optimal value,this paper introduces an improved dissipative particle swarm optimization algorithm to calculate the model coefficients.The improved dissipative particle swarm optimization algorithm can not only avoid the model falling into the local optimal value,but also maintain a high dissipation rate and increase the operating efficiency.This paper uses English phonemes and words as experimental sample data to establish a second-order Volterra speech prediction model based on the dynamic uniform search particle swarm optimization algorithm.(3)In order to reduce the model complexity,the key items of Volterra model are extracted in the scope of error allowed,which helps to reduce the number of parameters.The analysis of experimental results demonstrates that the chaotic speech time series prediction based on the combination of the hidden space reconstruction model and the improved dissipative particle swarm optimization algorithm for the single-frame and multi-frame prediction of the speech signal is superior to the linear model.Moreover,the accuracy and predictive performance of the algorithm are further improved compared with the basic dissipative particle swarm optimization and the Volterra model.It also has a further improvement in speech prediction waveform,speech quality and speech evaluation,which can meet the requirements of speech prediction.In this thesis we put forward a new method that combines the improved model of chaotic speech signal with the improved dissipative particle swarm optimization(DPSO),which introduces a new point of view for the speech signal prediction and it can be used for complex speech time series prediction.