This paper introduces the Voice Pre-processing and Quality Assessment Dataset (VPQAD), a scalable resource developed to validate various pre-processing techniques and improve voice signal quality in noisy environments. The dataset comprises voice recordings from 50 participants aged 18 to 40, captured in controlled real-life conditions using Audio Technica AT2020 and SHURE SM58 microphones. These high-quality recordings, made under diverse noise levels and settings, could be used for testing and developing voice enhancement algorithms. The dataset includes detailed metadata on the environment and participant demographics for analyzing and improving speech clarity and intelligibility, particularly in challenging conditions. To protect privacy, all data have been anonymized. VPQAD has been made public to promote collaborative research and advance research in biometrics, telecommunications, assistive technologies, and other applications requiring clear voice communication.

Self-supervised learning (SSL) models have achieved remarkable success in speaker verification tasks, yet their robustness to real-world audio degradation remains insufficiently characterized. This study presents a comprehensive analysis of how audio quality degradation affects three prominent SSL-based speaker verification systems (WavLM, Wav2Vec2, and HuBERT) across three diverse datasets: TIMIT, CHiME-6, and Common Voice. We systematically applied 21 degradation conditions spanning noise contamination (SNR levels from 0 to 20 dB), reverberation (RT60 from 0.3 to 1.0 s), and codec compression (various bit rates), then measured both objective audio quality metrics (PESQ, STOI, SNR, SegSNR, fwSNRseg, jitter, shimmer, HNR) and speaker verification performance metrics (EER, AUC-ROC, d-prime, minDCF). At the condition level, multiple regression with all eight quality metrics explained up to 80% of the variance in minDCF for HuBERT and 78% for WavLM, but only 35% for Wav2Vec2; EER predictability was lower (69%, 67%, and 28%, respectively). PESQ was the strongest single predictor for WavLM and HuBERT, while Shimmer showed the highest single-metric correlation for Wav2Vec2; fwSNRseg yielded the top single-metric R² for WavLM, and PESQ for HuBERT and Wav2Vec2 (with much smaller gains for Wav2Vec2). WavLM and HuBERT exhibited more predictable quality-performance relationships compared to Wav2Vec2. These findings establish quantitative relationships between measurable audio quality and speaker verification accuracy at the condition level, though substantial within-condition variability limits utterance-level prediction accuracy.

This paper presents the Extended-Length Audio Dataset for Synthetic Voice Detection and Speaker Recognition (ELAD-SVDSR), a comprehensive resource designed to advance both the generation of high-fidelity synthetic voices and the development of robust detection systems. Unlike existing datasets that typically provide short utterances, single-microphone recordings, or lack synthetic samples, ELAD-SVDSR uniquely integrates three critical features: 45-minute audio recordings from 36 participants, each reading various newspaper articles recorded under controlled conditions and captured via five microphones of differing quality, along with included deepfake voices for direct comparison. By having an extended duration of audio, ELAD-SVDSR captures a richer range of speech attributes such as pitch contours, intonation patterns, and nuanced delivery. This enables models to generate more realistic and coherent synthetic voices. The controlled microphone diversity enables systematic cross-device analysis, while the integrated synthetic samples eliminate the need for independent deepfake generation prior to detection research. Comparative evaluation demonstrates that the extended-duration training data may yield higher-quality deepfakes than those generated from existing datasets. Anonymized metadata accompanies the dataset to provide essential demographic information while preserving participant privacy. This dataset is anticipated to advance research in forensics, synthetic audio detection, biometric security, and voice authentication systems.

Field-programmable gate arrays (FPGAs) are increasingly used to bring biometric recognition from cloud- or GPU-centric deployments to resource-constrained edge devices where latency, power, and privacy are critical. This paper surveys recent (2021–2025) FPGA and FPGA-SoC implementations across five widely deployed modalities: face, fingerprint, iris, speaker (voiceprint), and finger vein. For each modality, we summarize representative implementations and the performance figures commonly reported in the literature (e.g., accuracy or EER, latency/throughput, resource usage, and power), highlighting the algorithm–hardware co-design choices that enable real-time operation. Across modalities, successful designs repeatedly employ streaming/dataflow architectures, aggressive quantization and fixed-point arithmetic, reuse-aware buffering, and heterogeneous CPU–FPGA partitioning, often supported by high-level synthesis and vendor deep learning IP. Beyond throughput, we discuss how FPGAs facilitate privacy-preserving on-device processing and can integrate template protection and presentation attack detection within the same fabric. Finally, we identify open challenges related to scalability to larger models, memory-bandwidth constraints, and design productivity, and outline research directions enabled by emerging adaptive FPGA architectures and more automated toolflows. Overall, the surveyed evidence indicates that FPGAs are a compelling platform for deterministic, energy-efficient, and secure biometric inference at the sensor edge.

Speaker recognition systems are increasingly deployed in security, forensics, and authentication, yet no standardized metric exists to predict the quality of a voice sample for recognition purposes. Existing audio quality metrics such as the Perceptual Evaluation of Speech Quality (PESQ) and Perceptual Objective Listening Quality Prediction (POLQA) measure perceptual quality rather than recognition utility, defined as a sample’s ability to produce accurate recognition outcomes.We propose the Voice Quality Index (VQI), the first utility-based voice quality metric designed to predict speaker verification performance. VQI employs a dual-score architecture producing two complementary integer scores on a 0–100 scale: VQIS for signal quality and VQI-V for voice distinctiveness. We extract 544 signal-quality and 161 voice-distinctiveness candidate features from each audio sample, selecting a subset of features through a three-stage pipeline. Random Forest classifiers trained on 20,288 balanced samples drawn from 1.2 million utterances across eight datasets produce utility-based quality scores. Evaluation using Error vs. Reject Curves on three test datasets (VoxCeleb1, VCTK, CN-Celeb) with five speaker verification providers demonstrates that rejecting the lowest-quality 20% of comparisons reduces the false non-match rate (FNMR) by up to 38.0% at the 1% FNMR threshold. A Principal Component Analysis (PCA)-based dimensionality reduction pipeline (77% for signal quality, 65% for voice distinctiveness) enables efficient deployment with minimal accuracy loss. VQI generalizes across English datasets and partially across languages, with quality-ordered partitions consistently separating system performance. An open-source desktop application is provided for practical deployment.

This paper introduces the Clarkson University Affective Dataset (CUADS), a multimodal affective dataset designed to assist in machine learning model development for automated emotion recognition, and the Affective Research Dataset Toolkit (ARDT), which contains both a mobile application to facilitate data collection and a Python API providing affective dataset abstractions to simplify data analysis and machine learning implementations. CUADS provides electrocardiogram, photoplethysmogram, and galvanic skin response data from 38 participants, captured under controlled conditions using Shimmer3 ECG and PPG sensors. Each participant used the ARDT mobile application to view and rate 20 movie clips. CUADS also provides Big Five personality traits for each participant. All data have been anonymized to protect participant privacy. CUADS and ARDT have been made public to promote collaborative research and advance research in automated emotion recognition and other applications requiring physiological signal processing.