Spoofing Detection Goes Noisy: An Analysis of Synthetic Speech Detection in the Presence of Additive Noise

Speaker recognition technology is recently finding its way to commercial applications to secure access to personal data or smart services. But similar to other biometric identifiers, speaker recognition is highly vulnerable to spoofing attacks where an attacker masquerades as a specific target speaker via text-to-speech (TTS) or voice conversion (VC) techniques in order to gain illegitimate access to the system. Existing studies on spoofing detection, determining whether a speech signal is natural or generated by a speech synthesis/voice conversion technique, mostly report their observations using high-quality clean recordings while the accuracy of spoofing detection in the presence of noise is largely unknown. To this end, we investigate the suitability of different synthetic speech detectors under additive noise. In order to find the most appropriate countermeasure, performances of various recently developed front-ends for spoofing detection are systematically examined under noisy conditions. We evaluate two different back-ends: Gaussian mixture model (GMM) and modern i-vector probabilistic linear discriminant analysis (PLDA) approach. Our experiments conducted on the recently released ASVspoof 2015 corpus indicate that synthetic speech detection accuracy decreases dramatically in the presence of additive noise. In contrast to the case of speech and speaker recognition system in noisy environment, we observe that standard speech enhancement techniques are not helpful. In the experiments with different front-ends, relative phase shift (RPS) features are found to give the highest synthetic speech detection accuracy for clean data on development set. However, Mel-frequency cepstral coefficients (MFCCs) and subband spectral centroid magnitude coefficients (SCMCs) are the absolute winners among the other features for matched and mismatched noise conditions, respectively.