关键词: 说话人识别, 录音回放攻击检测, 信道模式噪声, 统计帧, Legendre 多项式

Abstract:

The high fidelity of recent recording and playback devices with high quality and low price make it difficult to distinguish the playback recording from the authentic speech even by auditory perception,thus causing the playback attack to pose a huge threat to speaker recognition systems. In order to solve this problem,this paper proposes a playback attack detection method based on channel pattern noise. In this method,according to the channel
difference between the playback recording and the authentic speech,different channel pattern noises are introduced respectively for recording and playback devices in different channels and are then extracted by using the denoising filter and the statistical frame. Moreover,6-order Legendre coefficients and six statistical features are extracted from the channel pattern noise,and the support vector machine is used to establish the channel noise model to judge whether the input speech is a playback attack. Experimental results show that,by introducing the playback detector,the equal error rate of the speaker recognition system against the playback attack decreases by about 30%.

Key words: speaker recognition, playback attack detection, channel pattern noise, statistical frame, Legendre polynomial