Ashish Mani

Audio deepfake detectors that reach near-perfect accuracy on one benchmark dataset degrade sharply when evaluated on a corpus with speech synthesized by different methods, a failure attributed to memorization of corpus-specific synthesis artifacts. A frame-level separability analysis across two corpora with contrasting synthesis paradigms finds their artifacts almost entirely disjoint in static spectral content (of 28 tested features, only one is discriminative in both). Any transferable signature must therefore reside in temporal structure rather than in individual frames, motivating reservoir computing in which a fixed, untrained dynamical system supplies temporal abstraction that cannot itself adapt to the training corpus. We develop two such backbones: a 7-qubit transverse-field Ising quantum reservoir calibrated to its ergodic-MBL critical point, extended by a data re-uploading scheme that admits 60-dimensional gammatone features without enlarging the register; and a classical echo state network, compared under matched conditions. Both are trained under a novel asymmetric real-anchoring loss that scores utterances by distance from a shared real-speech centroid rather than a corpus-specific spoof signature. Under strict zero-shot cross-corpus evaluation, the two backbones specialize rather than one dominating. The ESN is stronger on both within-corpus conditions, reaching 3.27% EER on its best cell. The quantum reservoir is stronger on both cross-corpus conditions, reaching 24.61% EER compared to the classical reservoir's 42.64% EER. The quantum reservoir reverses the default cross-corpus performance ordering as well. Three independent attempts to unify both backbones into one dominant architecture were tried but yielded no further gains.

Salp Swarm Optimization (SSO) has attracted attention due to its simple structure and limited number of control parameters. However, when applied to high-dimensional problems or search spaces with strong inter-variable dependencies, the algorithm often experiences premature convergence and loss of diversity. To address this limitation, this work introduces Quantum-Entangled Salp Swarm Optimization (QESSO), an extended version of SSO that integrates quantum-inspired concepts, including Q-bit representation and entanglement-based interactions among decision variables. By allowing correlated updates between variables and incorporating a localized refinement strategy around the best solution, the proposed method enhances both exploration and exploitation capabilities. The performance of QESSO is assessed using the IEEE CEC 2017 benchmark suite, which includes unimodal, multimodal, hybrid, and composite test functions designed to emulate complex optimization landscapes. Comparative experiments against SSO, PSO, QPSO, and Differential Evolution (DE) indicate that QESSO consistently achieves improved convergence behavior, lower variance across runs, and better scalability as dimensionality increases. Although the evaluation in this study focuses on benchmark functions, the characteristics of these problems resemble those encountered in engineering design, control optimization, logistics planning, and other industrial applications. Direct validation on domain-specific industrial case studies is planned for future investigation.

This study investigates a novel Quantum-inspired Evolutionary Algorithm (QiEA) for denoising phonocardiogram signals using a fractional-order digital differentiator FIR filter (FODD). The filter coefficients were optimised using the proposed QiEA with Adaptive Rotation Gate Operator (ARGO). The proposed algorithm accelerates convergence towards optimal solutions based on fitness feedback, improving filter optimisation while clamping rotation angles to maintain algorithm stability. The performance metrics utilised were Scale-Invariant Signal-to-Distortion Ratio (SI-SDR), Signal-to-Noise Ratio (SNR), Peak Signal-to-Noise Ratio (PSNR), Correlation Coefficient (CC) and Zero Crossing Rate (ZCR). The results demonstrate that the proposed algorithm achieved a 25% improvement in the SI-SDR metric in highly noisy environments for pathological PCG signals. Detailed comparative analyses were conducted for single and combined-noise models against current state-of-the-art Deep learning (DL) methods. It exhibited an increase in the SI-SDR metric in the presence of APGN, by 17% and 24% for single and combined-noise models, respectively. In the presence of AWGN, the output increased by 9% and 11%, respectively, for both models in highly noisy environments. To quantify suppression of noise, ZCR was evaluated for the PCG signals before and after denoising. ZCR reduced by 98.4% revealing removal of high frequency noise artifacts for both pathological and non-pathological signals. The results position the proposed algorithm as a data-efficient, noise-adaptable alternative solution critical in resource-constrained dynamic situations across unseen noise patterns. These improved results can be attributed to reliance on evolutionary strategies rather than learned weights. The findings presented in this paper can have a significant impact on healthcare monitoring systems and real-time processing in critical systems.