基于AOF-LCNN的语音回放攻击场景下的说话人识别算法

Speaker recognition based on AOF-LCNN in audio playback attack scenario

针对语音回放攻击场景下的LCNN说话人识别系统中存在的过拟合问题,提出一种基于AOF-LCNN的神经网络。设计了一个新的DNN结构分类器作为后端分类网络,将该DNN结构级联在LCNN网络之后,形成一套新的端到端网络结构。由于LCNN结构中的MFM结构可能是造成过拟合的原因,在DNN后端结构中采用LeakyReLU作为激活函数,以抵消MFM的过拟合影响。在ASVspoof 2017数据集上的结果表明,该算法在Dev数据集和Eval数据集上分别达到了3.59%和13.79%的等错误率(EER),相对LCNN系统的等错误率分别降低了2.12%和3.51%。该算法一定程度上解决了过拟合的问题,提高了系统的鲁棒性,同时降低了系统的等错误率,从而提高识别性能。

Aiming at the over-fitting problem in LCNN speaker recognition system in audio playback attack scenario, a neural network based on AOF-LCNN is proposed. A new DNN structure classifier is designed as the back-end classification network and it is cascaded after the LCNN network to form a new end-to-end network structure. Because the MFM structure in the LCNN structure may be the cause of over-fitting, LeakyReLU is used as the activation function in the DNN back-end structure to offset the over-fitting effect of MFM. The results on the ASVspoof 2017 dataset show that the proposed method achieves an EER of 3.59% on the Dev dataset, an EER of 13.79% on the Eval dataset. The EER of the proposed method compared to that of the LCNN system was reduced by 2.12% and 3.51%, respectively. The proposed method solves the over-fitting problem to some extent, improves the robustness of the system and reduces the equal error rate of the system, thus it improves the recognition performance.