Spoof Speech Detection Based on Extended Constant-Q Symmetric Subband Cepstrum Coefficients and Fused Features

With the improvement of voiceprint synthesis and recording technologies, the malicious use of spoof speech will pose a significant threat to the automatic speaker authentication system. Therefore, it is essential to study speech feature extraction and spoof detection (SD). In this paper, some feature extraction methods for spoof speech are proposed. In the fore-end feature extraction, constant Q symmetric-subband cepstrum coefficients (CQSCC) and extended constant Q symmetric subband cepstrum coefficients (eCQSCC) based on phase symbol magnitude phase spectrum (PMPS) are proposed. In the back-end detection system, the features proposed in this paper are modelled through the Gaussian mixture model (GMM) and evaluated with EER and t-DCF. Under the physical access (PA) and logical access (LA) scenarios, the eCQSCC is optimized by 35.1% and 34.6%, respectively, compared with baseline feature CQCC. Further, the eCQSCC proposed above and prosodic features in other domains are fused at a scoring level and better anti-deception performance than features in a single domain is obtained.