基于听感知特征的语种识别

为了在语种识别时充分利用人的听感知特性提高识别性能,提出了一种基于听感知模型的特征。听感知特征采用Gammatone滤波器组代替常用的三角滤波器组计算语音信号各子带能量;根据等效矩形带宽模型,确定各滤波器的中心频率与带宽;使用反置等响度曲线模拟人耳对信号不同频率成分的主观响度感受。在基本听感知特征的基础上,还提出了一、二阶差分特征和偏移差分特征用于语种识别。对比实验表明,该文所提的听感知特征性能均优于目前普遍使用的Mel频率倒谱系数(MFCC)特征及其衍生特征。

Language identification based on auditory features

An auditory-based feature extraction algorithm was developed to improve the recognition performance of language identification algorithms using human auditory characteristics.The sub-band energies of the extracted auditory features were calculated using a Gammatone filter bank instead of the commonly used triangle filter bank.The center frequencies and bandwidths were then determined according to the equivalent rectangular bandwidth(ERB) model.The subjective human loudness perception for different frequency components was simulated by an inverse equal loudness curve.The first- and second-order delta cepstrum and the shifted delta cepstrum were derived based on these auditory features.Tests show that the features outperform the widely used Mel-frequency cepstrum coefficient(MFCC) counterparts.