Search Results (743)

The intersection of artificial intelligence (AI) and gamification in language education has attracted increasing attention, but its link to sustainability is still largely unexamined. This study looks at whether and how much this area of research tackles two aspects of sustainability: the ability of AI-driven gamified tools promoting lasting independent language learning (pedagogical sustainability) and their potential to reduce educational inequalities in line with Sustainable Development Goal 4 (SDG 4) (social sustainability). Using a sequential explanatory mixed-methods design, a bibliometric analysis of 105 documents retrieved from Scopus and Web of Science (WoS) (2018–2026), processed with bibliometrix R package (4.6.0) and VOSviewer (1.6.20), and a thorough qualitative content analysis of 27 studies were combined. The bibliometric mapping uncovered four thematic clusters around gamification design, motivational theory, serious games, and AI-driven vocabulary learning. No sustainability-related terms had enough density to form a clear cluster. The qualitative analysis showed that SDG 4 is missing from all 27 reviewed documents. Only three (11.1%) operationalizing pedagogical sustainability and mediation as a CEFR competence are absent. These findings suggest that sustainability is still on the fringe of this field and support a research agenda focusing on long-term measurement, self-directed learning support, low-resource design and mediation in AI-driven gamified environments. Full article

Background: Commercial digital mental health platforms have expanded across the Gulf since 2020, yet cancellation in this delivery model remains understudied. We aimed to identify factors independently associated with session cancellation on a Saudi Arabic-language commercial platform and to determine whether they describe user or provider behaviour. Methods: We conducted a retrospective cohort analysis of 36,544 booked sessions from 21,904 unique users (October 2023–April 2026); the outcome was binary cancellation (cancelled vs. completed). Multivariable logistic regression with cluster-robust standard errors estimated adjusted odds ratios (aORs) for user, provider, and booking characteristics; secondary analyses decomposed cancellation by recorded initiator and modelled age, lead time, and calendar time flexibly. Reporting followed STROBE guidelines. Results: Crude cancellation was 25.6%. Cancellation was strongly patterned by visit type (follow-up aOR 0.26, 95% CI 0.21–0.32), lead time (≥8 days 2.10, 95% CI 1.90–2.33), session type (free 1.19, 95% CI 1.10–1.27), and provider specialty; age reduced odds modestly (aOR 0.90 per decade, p < 0.001) and gender was null (aOR 0.95, p = 0.119). Cause-specific models showed that the lead-time gradient reverses by initiator: at ≥8 days versus same-day, user-initiated cancellation was less likely (aOR 0.62, 95% CI 0.49–0.78), whereas provider-initiated cancellation was far more likely (aOR 4.75, 95% CI 4.23–5.35), and provider-specialty differences were confined almost entirely to provider-initiated cancellation. The very low adjusted odds for emergency sessions (aOR 0.17, 95% CI 0.13–0.23) reflect how emergency bookings are administratively coded on the platform rather than genuine behavioural non-adherence and are regarded as artefactual. Discrimination was moderate (apparent AUC 0.700, development sample). Conclusions: Cancellation reflected operational features of the booking, but its largest correlates—lead time, free-session status, and provider specialty—operate predominantly through provider-side scheduling rather than patient non-adherence. Full article

Pixel-level annotation of remote sensing imagery is costly, motivating weakly supervised semantic segmentation (WSSS) using only image-level labels. However, class activation maps (CAMs) often highlight only discriminative sub-regions and fail to separate adjacent land-cover regions, particularly in remote sensing scenes characterized by densely co-occurring land-cover classes and substantial variations in object scale. To address these limitations, we propose a three-stage framework that integrates complementary priors from Contrastive Language–Image Pre-training (CLIP), Self-Distillation with No Labels version 2 (DINOv2), and the Segment Anything Model (SAM). First, a lightweight CLIP adapter aligns vision–language priors with remote sensing imagery, while sigmoid-based multi-label decoupled distillation replaces class-competitive distillation with independent class-wise supervision, producing more complete CAMs. Second, DINOv2-guided feature clustering decomposes large merged regions before SAM prompt generation, while Spatial–Semantic Constraints are used to construct confidence-guided point-and-box prompts and reject excessively expanded or semantically inconsistent masks, thereby generating reliable pseudo-labels. Finally, a compact segmentation network is initialized with the weights learned in Stage 1 and retrained using the refined pseudo-labels generated in Stage 2, eliminating the need for foundation models during inference. Experiments on the Potsdam, LoveDA, and DeepGlobe datasets show that the proposed method achieves mean intersection over union (mIoU) scores of 53.16%, 52.66%, and 62.98%, respectively, outperforming state-of-the-art WSSS baselines by 6.55, 1.16, and 1.27 percentage points, respectively. These results demonstrate the effectiveness and generalizability of the proposed framework across diverse remote sensing scenarios under image-level supervision. Full article

Background/Objectives: Acne vulgaris can impair quality of life (QoL) beyond lesion burden, and distress, insomnia, and stress may co-occur. We compared patient-reported burden across treatment-intensity groups used as pragmatic proxies of disease burden and management complexity, quantified associations between clinical severity and psychosocial measures, and modeled predictors of high QoL impairment. Methods: Cross-sectional outpatient study (N = 97) in Timișoara, Romania. Severity was assessed using the Global Acne Grading System (GAGS) and scarring grade (0–3). Participants completed Romanian-language versions of the Dermatology Life Quality Index (DLQI), Cardiff Acne Disability Index (CADI), Patient Health Questionnaire-9 (PHQ-9), Insomnia Severity Index (ISI), and Perceived Stress Scale-10 (PSS-10). Group comparisons used analysis of variance (ANOVA)/Kruskal–Wallis and chi-square tests; associations used Spearman correlations. Multivariable ordinary least squares (OLS) regression modeled DLQI; penalized logistic regression modeled DLQI ≥ 11. Treatment intensity was analyzed as a descriptive stratification variable rather than as an independent exposure, and a prespecified sensitivity analysis re-examined the DLQI gradient after adjustment for GAGS and scarring grade. The DLQI ≥ 11 cut-off was prespecified from validated DLQI banding. Clustering, mediation, and network analyses were exploratory and hypothesis-generating in view of the modest sample size. Results: DLQI differed across treatment intensity strata (topical only: 11.1 ± 5.7, oral antibiotic + topical: 16.9 ± 7.6, isotretinoin: 20.4 ± 7.1; p < 0.001). High impairment (DLQI ≥ 11) occurred in topical only: 40.6%, oral antibiotic + topical: 78.8%, isotretinoin: 87.5% (overall 69.1%; p < 0.001). DLQI correlated with PHQ-9 (ρ = 0.59), ISI (ρ = 0.49), and PSS-10 (ρ = 0.60) (all p < 0.001). In multivariable analysis, GAGS (B 0.35, p < 0.001), PHQ-9 (B 0.27, p = 0.017), and PSS-10 (B 0.24, p = 0.007) independently predicted DLQI. The combined predictive model achieved an area under the curve (AUC) of 0.841 (5-fold cross-validation, CV) for identifying DLQI ≥ 11. Conclusions: In this Romanian outpatient cohort, psychosocial measures explained substantial variability in acne-related disability alongside clinical severity. These cross-sectional findings support brief integrated screening and suggest that higher-burden treatment strata merit particular psychosocial attention, but they should not be interpreted as evidence that treatment intensity itself independently causes worse mental health outcomes. The contribution of this work lies less in confirming known associations than in quantifying, in an under-represented eastern European outpatient setting, how much of the identification of high-impact patients depends on patient-reported rather than lesion-based information. Full article

Camera-reconstructed 3D scenes often require offline visual cleanup before inspection, presentation, or reuse as renderable virtual-scene assets. Representative applications include removing temporary furniture, parked vehicles, equipment, signage, and other distracting or obsolete objects from reconstructed indoor and outdoor environments. Such editing requires not only accurate target localization across viewpoints but also plausible recovery of the previously occluded background. Existing methods often depend on manually specified masks or category-restricted detectors, while projection-based pipelines independently inpaint multiple views and subsequently refine the 3D representation, potentially introducing cross-view appearance and geometry inconsistencies. We present OVR-GS (Open-Vocabulary Removal in Gaussian Splatting), an instruction-driven object-removal framework for pre-trained 3D Gaussian Splatting (3DGS) scenes. Given a free-form instruction, a language parser generates target-oriented queries and a textual background-completion condition. Grounding DINO and the Segment Anything Model (SAM) produce multi-view candidate masks, which are filtered using Contrastive Language–Image Pre-training (CLIP). The proposed Semantic-Aware Gaussian Selector (SAGS) aggregates rendering-contribution-weighted mask evidence, groups spatially coherent candidates, and identifies the target Gaussian subset through rendered-cluster semantic verification. After removal, new Gaussians are initialized from boundary-adjacent primitives and interior samples and optimized locally using Score Distillation Sampling (SDS), while the original background remains fixed. On IMFine, SPIn-NeRF, and Inpaint360GS, OVR-GS achieves peak signal-to-noise ratio (PSNR) values of 19.78, 17.82, and 24.62 dB and Fréchet inception distance (FID) values of 142.30, 148.60, and 34.80, respectively. The results demonstrate the effectiveness of localized Gaussian optimization for instruction-driven cleanup of reconstructed environments before visual inspection, presentation, or reuse as renderable virtual-scene assets. Full article

Objective: Depressive symptoms exist on a continuum in the general population, yet the underlying neurobiological mechanisms, particularly the interplay between resting-state networks and task-evoked social cognitive responses, remain elusive. Methods: Leveraging the Human Connectome Project (HCP) dataset, we included 867 participants. With depression scores as the independent variable and age/sex as covariates, we systematically examined associations with sleep quality, negative emotions, sensory scores, gray matter volume (GMV), fractional amplitude of low-frequency fluctuations (fALFF), multi-seed resting-state functional connectivity (rsFC), as well as brain activation and behavioral performance during working memory, emotion recognition, social cognition, relational reasoning, language comprehension, and gambling tasks. The statistical threshold was set at voxel-level p < 0.001 (uncorrected) combined with cluster-level FWE correction at p < 0.05. Results: (1) Depression scores were positively correlated with sleep disturbances, negative emotions (anger/fear), and pain. (2) In resting-state, depression scores negatively correlated with ventral striatum (VS)–cerebellum/parahippocampal gyrus/fusiform rsFC, yet positively correlated with pregenual anterior cingulate cortex (preACC)–supplementary motor area (SMA) rsFC. (3) In task-fMRI, only the social task showed a positive association with task accuracy and regional activation in bilateral pre/postcentral gyri, superior temporal gyri, left middle frontal gyrus, and SMA/paracentral lobule. Conclusions: Elevated depression scores are linked to a pattern that may reflect relative decoupling between reward and perceptual systems, along with enhanced connectivity in cognitive control circuits. Socially, high scorers exhibit a pattern suggestive of compensatory hypervigilance, accompanied by enhanced behavioral performance. This study provides multidimensional evidence for the dimensional neural representation of depressive symptoms. Full article

IT help desks at large organizations face a high volume of recurrent, well-documented user requests that nevertheless require human-written replies, creating a persistent staff workload that is repetitive in content but non-trivial in tone and procedural correctness. We present FroLineR, short for Front-Line Response, a system that drafts the initial staff reply to such tickets in the login and account-activation category and integrates into a human-in-the-loop ticketing workflow on a Romanian-language ticketing platform. The generator is an unmodified instruct model augmented with retrieval from a small set of hand-curated guide documents, using a Romanian system prompt refined over several rounds of staff review. To evaluate and refine the prompt without manual labeling, we cluster the first user message of every historical thread with both BERTopic and Semantic Signal Separation ($S^3$), score configurations along coherence and lexical-diversity axes, and extract a 200-message evaluation set from the winning model. Prompt convergence was certified by several rounds of manual review by support staff. The production system is quantized to Q4_K_M GGUF, served through llama-cpp-python behind a small Flask API, and deployed with GPU offloading on the target server, reducing end-to-end per-answer latency from approximately 830 s on the server’s CPU to roughly 61 s once layers are offloaded to the GPU, with no observable degradation in answer quality. Full article

Background/Objectives: Dental caries remains one of the most prevalent and preventable childhood diseases worldwide. Despite universal access to free dental care in Saudi Arabia, preventable oral diseases remain common, highlighting persistent gaps in oral health literacy and preventive behaviours among children. Male students consistently demonstrate poorer oral health knowledge and practices than their female counterparts, yet they remain understudied. This study aimed to assess oral health knowledge, attitudes, and practices (KAP) among male primary school students in Al Lith city, Saudi Arabia, and to identify their associated sociodemographic determinants. Methods: A school-based analytical cross-sectional study was conducted during the 2023–2024 academic year in Al Lith city, Makkah Province, Saudi Arabia. A total of 430 male students aged 10–13 years from grades five and six were recruited via multistage cluster sampling from 8 randomly selected schools. A validated, interviewer-administered, Arabic-language questionnaire assessed knowledge (18 items), attitudes (12 Likert-scale items), and practices (12 items) related to oral health, alongside sociodemographic data. The study was designed and reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement. Data were analysed using IBM SPSS Statistics version 27. Descriptive statistics and multivariable logistic regression were performed. Results: Of the 430 participants, 83.3% were aged 11–12 years, 89.8% were Saudi nationals, and 89.1% resided in intact, two-parent households. Knowledge deficits were widespread: only 8.4% correctly identified the recommended timing of sweet consumption, 20.2% understood dental plaque, and 28.6% recognised the protective role of fluoride. Attitudes were generally positive, with 80.1% agreeing that teeth should be brushed twice daily and 71.6% endorsing regular dental visits; however, 60.3% held that dental attendance was necessary only when pain was present. Substantial practice gaps were also evident: only 33.3% brushed for the recommended two minutes, and 12.6% had never visited a dentist. In multivariable analysis, low household income was independently associated with better oral health knowledge (adjusted odds ratio [AOR] = 2.22; 95% confidence interval [CI]: 1.04–4.74; p = 0.039), divorced parental status was independently associated with positive oral health attitudes (AOR = 5.40; 95% CI: 1.28–22.79; p = 0.022), and older age was independently associated with poorer oral health practices (AOR = 0.72; 95% CI: 0.55–0.93; p = 0.014). Conclusions: Male primary school students in Al Lith demonstrated generally positive oral health attitudes but significant deficits in knowledge and preventive practices. Comprehensive, school-based oral health promotion programmes that address key misconceptions, actively engage parents, and integrate culturally accepted practices—such as miswak use—with evidence-based preventive guidance are needed. Full article

Background: The health sector is getting transformed with the usage of AI, be it diagnosis, treatment planning, disease prediction, and or health system management. Research in this field has picked up in the last few years, which was made possible with the emergence of machine learning, natural language processing and the increasing number of e-health records. Objectives: The study aims to investigate the current trends in the implementation of artificial intelligence (AI) applications in medical settings by investigating the global scientific output/landscape on this theme, such as the annual publication trends, country-wise contributions, and publishing patterns. Methods: The current study is based on systematic review by combining bibliometric analysis and cluster analysis using VOSviewer version 1.6.20, R software version 4.5.0, and Biblioshiny (Bibliometrix package in R) along with preferred reporting items for systematic reviews and meta analyses (PRISMA), 2020 which provides transparency and rigorous visualization to examine the articles published in English on the use of AI in healthcare, after the onset of COVID-19 till date i.e., from 2020 to 2026 on the Scopus database. Results: Using the relevant search string, 5940 documents were identified between 2020 and 2026, 1434 were included for analysis after screening and relevant filters. The publications have increased remarkably after 2020 on this theme and more than half of the publications have their roots in the discipline of Medicine. The USA, China, and the United Kingdom have contributed the most to the volume of research. Natural language processing and diagnosis are the emerging themes. The Journal of Medical Internet Research, BMC Medical Informatics and Decision Making, Computers in Biology and Medicine, IEEE Journal of Biomedical and Health Informatics, Frontiers in Public Health, and Digital Health are some of the most influential sources in the field. Li J and Liu X are among the authors with remarkable local impact. Conclusions: The work aims to assist investigators, health care professionals, and policymakers to learn about modern trends and focus on critical areas of future research and collaboration in AI-enhanced health care. The limitation of the study is that it considered only the Scopus database but it has opened up opportunities for researchers for analysis using other databases such as Dimensions, Lens, and PubMed. Also, this review is considering the publication record since the onset of COVID-19 but a comparative analysis of pre and post-pandemic studies can also be conducted to get a holistic view of drastic collaboration of research in this field. Discussions: The findings suggest that the role of artificial intelligence in health care has paramount over recent years, with other supporting technologies but a technologically hesitant population as well as low acceptance of AI due to ethical issues, cannot be ignored for ensuring efficiency in the health sector. Full article

Utah Lake and Great Salt Lake share a watershed yet face distinct pressures: eutrophication, harmful algal blooms, hydrologic decline, and exposed-playa dust hazards, but their combined literature has never been systematically characterized. I analyzed 1383 peer-reviewed records using bibliographic coupling, Louvain community detection, latent Dirichlet allocation, and large language model-assisted annotation. The coupling network (1012 records, 7536 edges) split into 379 communities ($Q=0.541$), of which 342 were singletons and only 14 reached 10 or more papers, together holding 59% of coupled records. Of 10 main-text clusters, 6 are anchored in a lake and 4 in a topic rather than a system; just 1 is Utah Lake dominant. Great Salt Lake research spans broader disciplinary communities (brine-shrimp ecology, mercury cycling, dust and paleoclimate); Utah Lake research is more applied (native-species management, eutrophication, harmful algal blooms). Recent Utah Lake studies report no lake-wide chlorophyll-a trend across 1068 Landsat scenes (1984–2021) and dissolved phosphorus near ∼0.02–0.04 mg L⁻¹ over ∼50 years despite ∼300% population growth, implying internal sediment cycling dominates; recent Great Salt Lake work centers on hydrologic decline and playa dust hazards. The synthesis identifies five cross-cutting gaps, including dust-emission-to-exposure assessment, Utah Lake nutrient modeling, and predictive water management, and yields a reproducible, transferable framework. Full article

The rapid adoption of large language models (LLMs) in financial services has generated a growing literature on “responsible AI” in domains such as investment analysis, credit assessment, risk management, compliance, and financial advisory systems. Unlike earlier AI systems, LLMs introduce responsibility challenges that differ in important ways from those addressed by earlier responsible AI frameworks, including hallucinations, prompt injection and manipulation, generative opacity, instruction-following failures, context sensitivity, output inconsistency, and emergent capabilities that arise only at larger model scales. While responsible AI concerns extend across AI systems more broadly, this review focuses specifically on LLMs, which since 2022 have become a major technological force reshaping financial services and have generated a distinctive body of responsibility discourse that existing frameworks are only beginning to address. However, the meaning of responsibility in this literature remains heterogeneous and inconsistently operationalized across studies. This paper provides a structured review of research on responsible LLMs in finance using a broad-to-narrow screening process across Web of Science and Scopus, reported with reference to the applicable PRISMA 2020 items. We combine a descriptive bibliometric overview with qualitative content analysis to examine the corpus. The descriptive overview covers publication trends, publication venues, disciplinary orientations, geographic distributions, and institutional affiliations, while the qualitative analysis codes the corpus on two dimensions: responsibility depth—whether responsibility is constitutive of the paper’s contribution, operative within its design, or peripheral to its framing—and responsibility mode—whether that engagement is normative, technical, evaluative, or mixed. The findings reveal a field concentrated in the Operative–Technical mode. Constitutive contributions cluster in Normative and Mixed modes, while a purely Evaluative orientation remains comparatively uncommon. Addressing these limitations requires moving beyond principle-level discourse toward institution-specific accountability mechanisms, empirical evaluation across clearly documented research and operational settings, and systematic alignment with relevant financial regulatory requirements. These efforts must address the distinctive challenges posed by LLMs, rather than only concerns inherited from the broader responsible AI literature. Full article

Accurate short-term building energy consumption forecasting is crucial for energy-system operational efficiency and demand-side management. Existing deep learning models heavily rely on historical data, making them costly for data-scarce buildings. Furthermore, their cross-building generalization ability is limited, requiring building-specific tuning that hinders large-scale deployment. To address these issues, this study proposes a novel framework based on large language models for short-term building energy consumption forecasting. It adapts large language models through parameter-efficient LoRA fine-tuning and incorporates building domain knowledge by using enhanced feature extraction modules and prompt design. Specifically, a prompt template rich in building physical semantics was designed to leverage the abundant pre-training knowledge of the large language model (LLM). This enables the model to avoid blindly fitting the data, directly aligning with building operating rules, providing a reasonable prediction basis even with limited data, and enhancing cross-building generalization. In addition, a cross-feature attention mechanism is designed to analyze the impact of dynamic meteorological features on energy consumption, thereby improving cross-climate scenario adaptability. Finally, to handle non-typical mutations in actual building operations, depthwise separable convolutional layers decompose residual components to filter out noise while preserving key features of anomalous occupancy patterns, thereby enhancing model robustness. Experiments on five real-world datasets have shown that the proposed framework outperforms state-of-the-art baselines, achieving an average improvement of 6.06% in the MAE and 4.80% in the RMSE. More importantly, it exhibits strong few-shot learning capabilities and extends to zero-shot forecasting. By reducing data dependency and enabling cross-building generalization, the framework developed in this work achieves scalable, low adaptation cost energy consumption forecasting from single buildings to clusters. Full article

Background/Objectives: In bilinguals, experience-based inter-individual differences, including language acquisition age, influence the functional and structural organization of the brain. Although many studies have explored the effects of age of acquisition on non-tonal languages, only a few studies have investigated the relationship between the age of tonal language acquisition and brain networks. Therefore, the aim of the study is to investigate how the age of language acquisition affects the topological organization of structural brain networks in proficient Japanese-Chinese/Chinese-Japanese speakers. Methods: A total of 33 proficient Japanese-Chinese/Chinese-Japanese speakers were recruited and divided into an early group into early and late groups according to their age of Chinese acquisition (AoCA). Sixteen of them constituted the early group, with AoCA at birth, whereas the late group consisted of 17 participants whose AoCA was after age 6. In addition to AoCA, the the age of second language acquisition (AoA-L2) was significantly younger in the early group than in the late group. All participants underwent diffusion MRI scanning. Structural brain networks were reconstructed, and graph-theoretical metrics were computed to compare topological properties between early and late groups. Results: Higher local efficiency and global clustering coefficient were found in the late group. For nodal parameters, the early group, compared to the late group, showed a higher nodal clustering coefficient in the right precuneus and higher betweenness centrality in the right Heschl’s gyrus (HG). Conclusions: Alterations in the right precuneus may be associated with earlier AoA-L2, consistent with its involvement in cognitive control processes required for managing multiple languages. Changes in the right HG could be related to earlier Chinese acquisition, particularly lexical tone learning, given the right HG’s role in pitch processing. This study showed how the linguistic environment, particularly language acquisition age, influences the topological properties of the structural connectivity, shedding light on its far-reaching effects on brain networks responsible for language processing and executive control. Full article

Crime offense descriptions are often recorded as unstructured text, making large-scale analysis and categorization difficult. This study proposes a semantic clustering framework for Thai crime offense descriptions using sentence embeddings, dimensionality reduction, and unsupervised clustering. Two datasets were obtained from Thonglor Metropolitan Police Station and Mueang Nonthaburi Police Station, Thailand. After preprocessing, the datasets contained 962 and 902 unique offense descriptions, respectively. Each description was transformed into a 768-dimensional embedding using SimCSE-PhayaThaiBERT. The embeddings were represented in Principal Component Analysis (PCA) Space and Uniform Manifold Approximation and Projection (UMAP) Space and clustered using K-Means, DBSCAN, HDBSCAN, and OPTICS. The results showed that UMAP Space generally provided more useful clustering results than PCA Space. Although DBSCAN achieved the highest internal clustering scores, it classified most records as noise. In contrast, HDBSCAN provided a more balanced result by maintaining strong clustering quality while retaining more records for interpretation. Qualitative analysis showed that the discovered clusters corresponded to meaningful offense categories. The proposed framework can support exploratory analysis of Thai crime records without requiring manually labeled data. Full article

The Greek dialects of Lesvos and Adrianople (present-day Edirne) during the time of the Ottoman Empire have much in common. Both are northern dialects and show considerable influence from Turkish; moreover, both were documented in the early 20th century, before the fall of the Ottoman Empire, by Kretschmer, in 1905, for Lesvos and by Ronzevalle, in 1911, for Adrianople. A comparative look at the two dialects reveals deep penetration of Turkish into their respective lexicons, with consequences for their phonology, and for Adrianople, into aspects of its grammar too. I focus here on a phonological matter, namely the conditions under which voiced stops (D) occur. Kretschmer’s song recordings from Lesvos show D only after full nasals (N), i.e., in ND clusters, while later documentation, Newton 1972, shows no nasality, with only D as the realization, even in words with ND earlier. Adrianople Greek shows intervocalic D but also ND clusters, only in Turkish loans and always matching the consonantism of the Turkish source; such is the case too in ND/D variation in rebetika songs from Istanbul. This suggests that the issue really involves a kind of dialect mixing, and it is argued that this holds even for Lesvos, and even for sociolinguistically coded variation. Thus, bringing Adrianople Greek into consideration clarifies the nature of voiced stop realizations in Lesvos phonology. Full article