A Contrastive Learning-Based Short Speech Bio-key Generation Model

With the widespread deployment of biometric authentication in critical security applications, bio-key generation technology has attracted increasing attention. The extraction of secure, stable, and high-entropy keys from biometric traits has become a central research challenge. However, existing bio-key generation methods fail to fully satisfy security requirements in practical scenarios, as the generated keys are vulnerable to attacks and may even lead to leakage of sensitive biometric information, resulting in unpredictable consequences. Moreover, variations in voiceprint acquisition conditions, noise interference, and the inherent instability of voiceprints make it difficult for current methods to achieve an effective balance between key consistency and security. To address these challenges , this paper proposes a short-speech bio-key generation model based on contrastive learning. The proposed model introduces a contrastive learning mechanism to align features across multiple samples of the same voiceprint while enhancing the discriminability between different voiceprints, thereby obtaining robust and highly discriminative voice representations. On this basis, feature quantization and error-correcting codes are employed to generate stable bio-key with high randomness. The proposed model consists of three main modules: the voiceprint information preprocessor, the voiceprint feature vector extrac-tor, and the voiceprint key fuzzy extractor. The experimental evaluation in this study was conducted on a self-collected dataset, namely GUIT VP01. All experimental results were obtained through comprehensive testing and performance validation on this dataset. Experimental results show that the proposed model generates keys with an entropy exceeding 1024 bits, an accuracy > 99%, and a misrecognition rate < 0.01%. These results demonstrate that the proposed voiceprint bio-key generation model effectively meets user requirements for both key strength and security.