结合谱减和缺失特征重建的鲁棒性话者识别

Missing Feature Reconstruction with Spectral Subtraction for Robust Speaker Identification

针对加性噪声影响下文本无关说话人识别系统性能急剧下降的问题，提出谱减和缺失特征重建相结合的方法。该方法将被噪声严重污染的频段称为缺失特征，采用谱减法对语音信号进行前端处理，并检测缺失特征；利用基于聚类的重建方法，由可靠特征估计缺失特征。在YOHO数据库上信噪比（SNR）为5～20dB的实验表明，相对于单独的谱减法和缺失特征重建方法，该方法的识别性能有显著提高。

Considering the problem of training/testing mismatch caused by the additive noise in the text-independent speaker identification framework, a new method for the combination of the spectral subtraction and the missing feature reconstruction is presented. The method labels highly corrupted features as missing features,adopts the spectral subtraction to mitigate the noise and detect missing features ,and uses the cluster-based reconstruction to recover missing features. Experimental results on YOHO data base show that the combined method outperforms the spectral subtraction and the missing feature reconstruction alone when SNR is 5- 20 dB.