基于神经元阈上非周期随机共振机制的语音复原技术研究

Speech Restoration Based on Suprathreshold Stochastic Resonance of Neuron Model

应用随机共振机制,通过噪声能量来加强语音信号,改善低信噪比语音的输出质量。对FitzHugh-Nagumo(FHN)神经元模型中存在的阈上非周期随机共振现象进行了分析,根据其阈值特性,此二维神经元模型可被等价为两状态的阈值跨越非线性动力学系统。因此对含噪语音信号添加噪声,产生阈值化后的二值输出,经迭代收敛进入阈上非周期随机共振状态。在一个非零添加噪声强度上,含噪语音输出的互相关系数将达到最大值。通过语音复原的结果表明,本文方法对噪声的变化有更好的鲁棒性,尤其在强背景噪声下,随机共振方法较其他传统方法有更佳的复原效果。

The quality of low signal-noise-ratio speech signal was improved by adding the noise energy to speech signal based on stochastic resonance mechanism. The suprathreshold stochastic resonance phenomenon of non-periodic response of FitzHugh-Nagumo （FHN） neuron model was analyzed firsdy. According to the threshold property, this two-dimensional neuron model may be equivalent to one nonlinear dynamic system, which has two states under the threshold condition. Adding the noise to the corrupted speech sig- nal, the two-value output was produced. The system was driven to the state of non-periodic suprathreshold stochastic resonance after the iterative convergence. The mutual relation ratio of the speech output will reach the maximum by adding a non-zero noise. The speech recovery results show that this method has a better robustness, particularly in strong background noise, stochastic resonance method than other conventional methods has a better recovery result.