| With the development of speech synthesis,voice conversion and the improvement of recording equipment quality,the traditional speaker verification system is increasingly threatened by various spoofing techniques.These threats can invade the speaker recognition system;therefore the ASV system will be interfered and the recognition accuracy will sharply decrease.Among all kinds of spoofing attempts,the threat of playback spoofing to the system is the most serious.In order to detect fraud in audio playback,an anti-spoofing method based on complementary features is proposed in this paper.The original audio contains abundant amplitude and phase information.Because the human ear is believed not sensitive to phase information and it is difficult to model and process phase information,the traditional feature extraction methods abandon phase information while feature extracting.In fact,there is phase wrapping while recording which could be well used to detect the spoof.Secondly,traditional feature extraction methods pay more attention to low-frequency information by imitating human ear auditory characteristics,while playback audio has a lot of high-resolution information in high-frequency band because of channel noise and other factors.Therefore,in this paper,for the fusion of phase and amplitude information,the fusion of high-frequency and low-frequency information is expected to improve the detection effect of playback fraud.This paper tries to compare the effects of traditional feature extraction methods in spoofing detection,and proposes complementary features of amplitude and phase information.The paper also proposes an adaptive filter based on F-ratio full-band analysis for the first time to maximize capturing high discriminative information of playback audio and original audio in frequency domain,and then proposes an adaptive phase and amplitude feature extraction method.The proposed method reduces the relative error rates of 78.7% and 59.8% respectively in the development set and the evaluation set of ASVspoof 2017 database. |
