听觉模型鲁棒性特征研究及应用

Research and Application of Robust Characteristics of A uditory Models

人类的听觉系统具有非常精细而巧妙的结构,即使在嘈杂的环境中,也能准确地理解语音。采用精细的耳蜗模型作为前端处理可以实现更好的语音处理。利用快速压缩的非对称谐振器级联(CARFAC)作为人耳外周模型,结合听觉稳定图像得到精确的皮层前听觉模型。在听觉模型的基础上提取较准确的基音轮廓,利用基音信息进行声场景分析,合成鲁棒性语音特征,并将其送入神经网络进行监督训练,以实现语音增强。实验结果表明,噪声条件下,由听觉模型提取的特征在各语音评价指标下都有较好的体现,可以更好表征语音信号,具有一定的鲁棒性。

The human auditory system has a very fine and ingenious structure,and it can accurately understand speech even in a noisy environment.Using a fine cochlea model as front-end processing allows for better speech processing.In this paper,a rapidly compressed asymmetric resonator cascade(CARFAC)is used as a peripheral model of the human ear,combined with an auditory stabilization image(SAI)to obtain an accurate precortical auditory model.Based on the auditory model,a more accurate pitch contour is extracted,the pitch information is used to analyze the acoustic scene,and robust speech features are synthesized,which are sent to the neural network for supervised training to achieve speech enhancement.Experiments show that under noise conditions,the features extracted by the auditory model are better reflected in various speech evaluation indicators,which can better characterize the speech signal and have certain robustness.