基于声道频谱参数的语种识别

Language Identification Based on Vocal Tract Spectrum Parameters

针对低信噪比下语种识别正确率低的问题,提出了一种声道冲激响应频谱参数和Teager能量算子倒谱参数融合的识别方法.根据语音中不同特征信息量分布特性,首先在特征提取前端引入低通滤波器滤除信号高频部分,并采用重采样方法降低采样率,再基于信号频谱提取声道冲激响应频谱参数,然后融合Teager能量算子倒谱参数,最后通过高斯混合通用背景模型进行语种识别验证.不同信噪比条件下性能测试表明,所提方法相对于基于单一的梅尔频率倒谱系数特征、单一的伽玛通频率倒谱系数特征和基于对数梅尔尺度滤波器组能量特征,在低信噪比下提升约15 dB,显著提高了识别正确率.

Aiming at the problem of low accuracy of language identification under low signal to noise ratio,a fusion identification method is proposed,using spectral parameters of channel impulse response and Teager energy operators cepstral coefficients.Considering the distribution of different feature information in speech,a low-pass filter is introduced to filter out the high-frequency part of the signal in the front-end of feature extraction.The resampling method is used to reduce the rate.And then,the spectral parameters of channel impulse response of vocal tract are extracted,and fused with the Teager energy operators cepstral coefficients.Finally,a Gaussian mixture model-universal background model is used to perform the language identification.Experiments under different signal to noise ratio conditions show that the proposed methold significantly improves the language identification accuracy with 15 dB gain at low signal to noise ratio compared with the single Mel frequency cepstrum coefficient feature,single Gammatone frequency cepstrum coefficient feature and log Mel-scale filter bank energies feature.