Knowledge-based authentication mechanisms such as passwords remain the dominant gatekeepers of digital systems, yet they continue to underpin the majority of credential-related data breaches due to their inherent vulnerability to theft, phishing, and reuse. Biometric authentication, particularly voice-based verification, offers a contactless and hardware-accessible alternative, but existing implementations are either confined to proprietary vendor ecosystems or require deep expertise in machine learning to deploy. This paper presents VoiceSecure AI, an open and modular voice biometric authentication framework that fuses classical signal-processing-based acoustic feature analysis with the contextual reasoning capability of large language models. The system extracts a 70dimensional acoustic feature vector composed of Mel-Frequency Cepstral Coefficients (MFCC), spectral descriptors, and prosodic measurements from a 16 kHz voice sample, transcribes speech using OpenAI Whisper, and submits both the features and transcript to GPT-4 for natural-language voice characteristic analysis and crosssample comparison. A 60/40 weighted fusion of cosine-based geometric similarity and the GPT-4-estimated samespeaker probability produces a final authentication confidence score, which is compared against configurable thresholds to render a verdict. All user data, voice profiles, and audit logs are persisted in a four-table SQLite schema. Empirical evaluation on a multi-session pilot dataset reports a True Acceptance Rate of 93.7%, a False Acceptance Rate of 2.4%, and an Equal Error Rate of approximately 4.1%, with an average end-to-end authentication latency of 4.6 seconds. Results indicate that LLM-augmented multi-factor fusion yields measurable accuracy and interpretability improvements over purely numerical baselines while remaining accessible on commodity hardware.
Sekhar et al. (2026) studied this question.