基于量子粒子群优化广义回归神经网络的语音转换方法

Voice conversion based on quantum particle swarm optimization of generalized regression neural network

针对粒子群算法优化神经网络进行语音转换时容易产生收敛速度慢、早熟的问题,本文采用一种新的量子粒子群算法优化广义回归神经网络的语音转换模型。该量子粒子群通过改变量子比特相位进而改变位置矢量,并利用量子非门进行变异操作。因此首先利用量子粒子群对网络进行优化得到最佳的光滑因子参数,从而建立频谱映射规则。接着,利用频谱参数和基频参数的相关性,对韵律特征基频也进行转换。然后,联立转换后的频谱参数和基频参数,利用STRAIGHT模型合成目标语音。最后,采用主观和客观测评方式进行评价。实验结果表明,与传统粒子群算法优化广义回归神经网络相比,该方法转换后的语音自然度和相似度得到提升,谱失真率下降2.1%。本文方法具有比径向基神经网络、广义回归神经网络、粒子群算法优化广义回归神经网络更好的语音转换性能。

In this paper,a new quantum particle swarm optimization algorithm is used to optimize the voice conversion model of generalized regression neural network in order to solve the problem of slow convergence and premature phenomenon in particle swarm optimization.The quantum particle swarm optimization algorithm changes the position vector by changing the quantum bit phase and uses the quantum non-gate to perform the mutation operation.Therefore,we first use the quantum particle swarm to optimize the network to get the best smooth factor parameters,so as to establish spectrummapping rules.After that, we use the correlation between the spectral parameters and the fundamental frequency parameters to convert the prosodic characteristic fundamental frequency.Then,the STRAIGHT model is used to synthesize the target voice in conjunction with the converted spectral parameters and the fundamental frequency parameters.Finally,we use the subjective and objective evaluation methods to evaluate.The experimental results show that the natural and similarity of the proposed method for the transformed voice are improved and the spectral distortion rate is reduced by 2.1% compared with the traditional particle swarm optimization algorithm.The proposed method has better voice conversion performance than radial basis function neural network,generalized regression neural network and generalized regression neural network optimized by particle swarm optimization.