隐相空间的DUPSO-RPSOVF语音预测模型研究

A type of nonlinear prediction model for speech signals based on second-order Volterra series is put forward. In order to overcome some intrinsic shortcomings caused by using the classic least mean square (LMS) algorithm to update Volterra model kernel coefficients, a dissipative uniform particle swarm optimization (DUPSO) algorithm is applied to obtain the kernel coefficients and then a DUPSO-SOVF prediction model can be constructed. A DUPSO-SOVF prediction model with hidden phase space is constructed by dynamically obtain parameters of embedding dimension and time delay in the process of solving model kernel coefficients rather than using traditional phase space reconstruction process. On the purpose to reduce model complexity, the key model kernels are extracted within the margin of the allowable error and the model kernels are then reduced, and the reduced parameter DUPSO-SOVF (RPSOVF) prediction model is proposed. Simulation results for samples of English phonemes, words and phrases show that:the DUPSO-SOVF model with hidden phase space can accurately calculate parameters of embedding dimension and delay time of phase space reconstruction; both of the DUPSO-SOVF model and the DUPSO-RPSOVF model exhibit higher prediction accuracy on single frame and multi-frame speech signal than PSO-SOVF and LMS-SOVF models. Also, the two proposed models can better reflect trends and regularities of the speech signal series and meet requirements for speech signal prediction.