Detection of speaker liveness with CNN isolated word ASR for verification systems

Abstract

The article proposes a new speaker liveness test for speech verification systems. Biometric authentication systems based on speaker verification are often subject to presentation attacks which use the target speaker's recorded speech. We propose a liveness test which uses CNN isolated word ASR as a countermeasure to repel attacks during the verification process. The liveness test incorporates the extraction of MFCC coefficients and the CNN classifier. Reliability of the recognition of isolated words is verified against a validation dataset of various sizes. The achieved results verified the system's reliability, which decreased slightly as the size of the keyword dataset increased. The proposed method represents a simple and effective security component against presentation attacks for existing SV systems.