用于室内环境说话人识别的混响补偿方法

A Reverberation Compensation Method for Speaker Recognition in Rooms

针对训练和识别环境不同而导致室内说话人识别系统识别率显著下降的问题,提出了一种基于差异化特征提取的混响补偿方法。与使用传统MFCC特征的识别阶段不同,该方法在训练阶段通过Schroeder反向积分在mel频带获得房间声能量衰减曲线,并使用该曲线补偿纯净信号的MFCC特征,以模拟实际室内混响场声信号特征;同时,通过联合应用相对谱滤波(RASTA)与倒谱均值规整(CMN)处理MFCC特征,进一步抑制房间通道效应对语音信号影响。针对不同混响程度房间中实测数据的识别结果表明,该方法可以显著提高识别率,具有良好的抑制混响作用。

To overcome the problem that the accuracy of speaker recognition systems in rooms descends rapidly as a result of the mismatch between training and testing environments, a differential feature extraction method based on reverberation compensation has been brought forward. Different from the recognition phase that uses traditional MFCCs, Schroeder inverse integration is applied to obtaining the energy decay curve in rooms, so that reverberation can be compensated for MFCC features of pure sound signals in training phase. Furthermore MFCCs are processed by CMN （ Cepstral Mean Normalization） and RASTA to suppress the room channel effect. The experimental results in different real rooms with various reverberation degrees and their analysis have shown preliminarily that the method we presented can enhance the recognition rate and performs well in suppressing the influence of reverberation.