#101: Zombie Interlocutors: How Hard Is Natural Language Processing?
Anders Søgaard
The hard problems of philosophy seem to have made their way into natural language processing (NLP) and computational linguistics. Is NLP solved with LLMs? If not, why not? The debate has led many of us to revisit classical debates of 17th century and early 20th century philosophy. I describe the philosophical landscape, as well as the relevant, empirical results of LLMs, and try to make ends meet.
#102: AI: The Good, the Bad, and the Ugly
Giuseppe Riccardi
Due to the rise of easy-to-use language-based systems, generative AI has taken center stage across many industries. Neophytes and practitioners in public and private organizations have been asked to work differently and to create new products and services by delegating and supervising token-crunching computer entities. Individuals are using these systems nonchalantly for personal advice (e.g., nutrition) or social fulfillment (e.g., companionship). In the meantime, AI researchers have unveiled critical limitations of the underlying machine learning models that could hinder their inclusion in innovation processes, products, and services. This talk will review the current state of the art in conversational AI, its severe limitations, and future challenges. Beneficial technology and systems must be created ab initio through an evaluation-first approach. We will give insights into how to create fundamental research challenges while pursuing innovation in AI system training and development.
#1371: "Chi nas dal soch el sent de legn" - Auditing Text Corpora for Lombard
Edoardo Signoroni and Pavel Rychlý
Several of the world's languages are still under-resourced in terms of Natural Language Processing (NLP) tools. This is mostly due to the lack of high-quality datasets to train, develop, and evaluate systems and models for several tasks, such as Machine Translation (MT). We conduct a manual audit of the parallel and monolingual corpora available for Lombard, an under-resourced language continuum from Italy. Our analysis reveals that the perceived abundance of web-scraped data is an illusion, with massive datasets plagued by severe language misidentification, boilerplate text, and non-linguistic noise. Furthermore, we analyze the orthographic composition of the valid Lombard portions across web-scraped datasets, curated corpora, and benchmarks. Our findings show conflicting orthographical systems and severe representational bias across all corpora: high-quality data is heavily skewed towards Western Lombard varieties, with Eastern ones left on the margins. This underscores the need for variety-aware, community-driven data curation rather than purely quantity-driven scraping.
#1409: Acoustic Volume Mismatch in Distant Streaming ASR: Case Study and Mitigation via Wide-Range Volume Perturbation
Martin Motejlek, Jiří Málek and Petr Červa
This paper focuses on acoustic volume mismatch between close-talk and distant speech, which degrades the performance of end-to-end (E2E) automatic speech recognition (ASR) systems. Our case study demonstrates that an E2E model trained on close-talk has a limited "effective volume range" and a performance bias, both mirroring the power distribution of its training set. We show that conventional mitigation methods, such as the narrow-range volume perturbation standard in Kaldi recipes, are insufficient to address this issue. Guided by these findings, we design a volume perturbation augmentation with a wide gain range on the decibel scale. Experiments confirm consistent improvements on distant speech with minimal degradation on close-talk, effective for both small and large training sets. Finally, we address rapid volume transitions, common in conversations between near and far speakers. We propose a proof-of-concept block-wise volume perturbation, opening an avenue for future research.
#1365: An Effective Structured Generation Approach to Sentence Stress Detection with Speech Language Models
Ting-An Hung, Yu-Hsuan Hsieh, Tien-Hong Lo, Hong-Yun Lin, Yung-Chang Hsu and Berlin Chen
Sentence stress marks the emphasis placed on particular words in speech and often conveys contrast, focus, or speaker intention beyond the literal transcript. Recent Speech language models enable direct audio-conditioned generation, making them naturally suited to alignment-free sentence stress detection (SSD). However, effective ways to formulate and adapt these models for robust stress detection remain underexplored, especially when reference transcripts are unavailable. To address this gap, this paper proposes a structured text generation framework that adapts Qwen3-ASR for transcript-free SSD. Instead of relying on forced alignment or an auxiliary classification head, the model generates a transcription in which stressed words are explicitly marked with predefined tags. We compare transcript-conditioned and direct stress-tagged generation, and further examine how acoustic conditioning and encoder adaptation affect generalization. Empirical findings show that direct stress-tagged generation generalizes better than transcript-conditioned generation, while freezing the audio encoder further improves robustness, reaching an F1-score of up to 0.9077 on out-of-domain evaluation and outperforming a strong WhiStress baseline in the transcript-free setting. Results suggest that direct structured generation, together with the preservation of pre-trained acoustic representations, provides an effective formulation for transcript-free sentence stress detection with Speech Language Models.
#1354: Analysis and Implications of Alignment Between LLM-as-a-Judge and Human Evaluators in a Generative AI Chatbot: A Real-World Study
David Escudero-Mancebo, Melquiades Rodríguez and Guillermo Vicente-Oliva
The emergence of generative artificial intelligence has simplified the development of conversational systems and improved interaction with chatbots, fostering their rapid adoption in user support services, including in the public administration sector. However, empirical evidence on how these systems should be evaluated in real-world contexts remains limited, particularly regarding the alignment between automated and human assessments. This study investigates the extent to which large language models (LLMs) can reliably approximate human judgment in the evaluation of chatbot responses. To this end, a generative AI–based chatbot designed to assist students during the university admission and enrollment process was developed using the Microsoft Copilot Studio platform. We compare human expert ratings with a document-grounded LLM-as-a-judge evaluation. Results show limited overall agreement but stronger alignment for factual errors, suggesting that automatic evaluation can support monitoring but not replace expert review. These findings contribute to a better understanding of the role of LLMs in evaluation pipelines and highlight both their potential and limitations for the reliable and responsible deployment of generative AI chatbots in public service environments.
#1383: Benchmarking Pragmatic Reasoning of Large Language Models in the Czech Language
Yevhenii Karpizenkov, Aleš Horák and Radim Lacina
Understanding how large language models recover meaning from context, beyond the literal content of an utterance, is an important part of evaluating their language abilities. Existing benchmarks for pragmatic reasoning remain focused mainly on English, leaving Czech and other morphologically rich or pragmatically flexible languages largely outside current evaluation. We introduce CZECH-PUB, a Czech benchmark for evaluating pragmatic understanding in large language models. The benchmark covers three pragmatic phenomena: presupposition, implicature, and information structure. It combines corpus-based extraction, controlled generation, and translation-based adaptation of selected conversational data in a unified multiple-choice format. The resulting dataset contains 1,920 items and allows both aggregate model comparison and detailed analysis of performance across the chosen pragmatic phenomena. We evaluate a range of current language models on CZECH-PUB and compare their performance across the three phenomena. The results show that performance depends strongly on both the model and the pragmatic phenomenon. While stronger models are relatively reliable on presupposition and implicature, information structure remains harder and more variable across systems. In particular, models vary substantially in how they interpret Czech word order as a discourse-sensitive cue.
#1394: Benchmarking Pretrained Speech Models for Arabic Dialect Identification with Cross-Corpus Evaluation
Christopher Simic, Korbinian Riedhammer, Sebastian Bayerl and Tobias Bocklet
Arabic dialect identification (ADI) is a challenging speech classification task due to the high phonetic and acoustic similarity between dialects and strong domain variability. While recent approaches report accuracy gains through increasingly large models, limited attention has been paid to how different pre-training strategies, model architectures, and data domains influence the quality and structure of learned representations. In this work, we present a systematic analysis of Arabic dialect identification focusing on representational effectiveness and generalization rather than accuracy alone. We compare ECAPA-TDNN models pre-trained on non-speech audio, speaker recognition, language identification, and dialect identification tasks, as well as transformer-based architectures including Whisper and wav2vec 2.0 at multiple model scales. All models are fine-tuned and evaluated on a curated Arabic dialect dataset covering five major dialect groups. In addition to in-domain evaluation, we conduct cross-corpus experiments between ADI17 and an internal dataset (ADI-int) to assess robustness under domain shift. Our results show that pre-training on large-scale language identification datasets consistently outperforms task-specific dialect pre-training, achieving up to 57% relative error reduction. Scaling transformer models improves performance, though with diminishing returns at larger sizes. To improve interpretability, we analyze embedding spaces using t-SNE projections combined with Kernel Density Estimation (KDE), revealing that stronger models learn more compact and better-separated dialect representations. Cross-corpus experiments expose substantial performance degradation, highlighting the limitations of fine-tuning for domain robustness. These findings emphasize the importance of diverse pre-training objectives and provide actionable insights into representation learning for Arabic dialect identification.
#1334: Blackwell: An Explainable Anamnesis-based Multi-Agent Framework for Remote Clinical Diagnosis
Lukas Torquato, Rejwanul Haque, Paul Stynes, Greg Doyle and Oisin Cawley
The adoption of Artificial Intelligence in telemedicine is currently constrained by the "Black Box" aspect of deep learning models and the lack of autonomous data acquisition mechanisms. This paper presents Blackwell, a Deep Agent Framework designed to automate the complete clinical pipeline---from patient interview to evidence-based treatment planning---while remaining explainable. Powered by LangGraph, LangSmith and Gemini, the system features a novel Anamnesis Agent that autonomously conducts safety-guarded patient interviews, feeding a Two-Phase Evaluator that separates diagnostic hypothesis generation from treatment research. On a 250-case synthetic dataset, Blackwell obtained 84% symptom recall, 86.8% emergency detection accuracy and 90.1% diagnostic accuracy, showing outstanding results for chronic and symptom-characteristic diseases. However, text-only diagnosis appeared to be insufficient for conditions based on visual inspections or serological markers (e.g. HIV/AIDS). Blackwell demonstrates that agentic AI can successfully bridge the gap to cold-start a remote clinical diagnosis with only patient conversation, combining a static knowledge base (RAG) with dynamic research (PubMed), offering a transparent, safe support tool for remote clinical assessment.
#1355: Building Data Infrastructure and LLM Benchmarks for the Masurian Dialect
Danuta Roszko, Roman Roszko and Piotr Szatkowski
The rapid evolution of Large Language Models (LLMs) widens the divide between high- and low-resource languages. We present a newly developed data infrastructure for Masurian -- a dialect characterised by non-standardised orthography and minimal digital presence. Our original four-stage pipeline transforms texts into raw, unified, normative, and translational datasets for machine learning. Developed within the CLARIN-PL framework following FAIR principles, this infrastructure supports the Masurian Linguistic and Cultural Competency Benchmark (Masurian L&CC). By utilising unique analogue data as a gold standard, the benchmark robustly evaluates LLMs' Natural Language Understanding (NLU), Natural Language Generation (NLG), and cultural competency, effectively addressing data contamination. This work is part of a broader initiative within the CLARIN-PL-BIZ-Bis project, aimed at preserving low-resource language heritage and advancing LLM development.
#1317: Communicative Function Classification using Voice Activity Projection
Michael Paierl and Barbara Schuppler
Spontaneous conversation involves complex turn-taking behavior that needs to be recognized in real time for natural human-robot interaction. A promising approach to this challenge is the Voice Activity Projection (VAP) model, which continuously predicts upcoming voice activity. However, it is trained primarily on relatively clean, English conversational speech, lacks evaluation on lively conversation and does not distinguish the communicative function of the current turn. We address these gaps using the Graz Corpus of Read and Spontaneous Speech, an Austrian German dataset with fine-grained communicative function labels, where zero-shot evaluation shows that VAP predictions do not generalize well. Hypothesizing that VAP embeddings nonetheless encode rich turn-structural information, we use them as input to a causal Long Short-Term Memory model to predict five communicative function classes (i.e., change, hold, hrt (backchannel), laughter, breathing) plus an absence class. We achieve a mean macro F1-score of 57.3% (SD 3.5%) across leave-one-out cross-validation for 13 speaker pairs in this six-class setup. Since hold and change represent the central turn-taking distinction and proved hardest to classify, we further conduct an acoustic analysis, finding that duration, intensity, and voice quality features strongly affect classification outcome. These findings highlight the challenges of fine-grained communicative function prediction in natural and spontaneous speech and provide a foundation for building real-time turn-taking components in human-robot interaction systems.
#1336: Comparing Acoustic and Linguistic Features for Automated Deception Detection
Jan Šimek, Marie Kunešová and Jindřich Matoušek
Humans are generally poor at recognizing lies, which makes potential deception detection systems beneficial for high-stakes environments such as legal trials, interrogations, and job interviews. This paper presents a comparative study of acoustic and linguistic features for automatic deception detection using the Columbia SRI Colorado (CSC) corpus. We establish acoustic baselines using a reduced set of 32 top-ranked features selected from the ComParE feature set alongside deep wav2vec 2.0 embeddings and evaluate their performance against two linguistic representations extracted from utterance transcriptions: frequency-based TF-IDF and contextualized Sentence-BERT (S-BERT) embeddings. A variety of machine learning algorithms were evaluated across single-modality and multimodal configurations. Our results demonstrate that acoustic features alone fail to reliably distinguish deceptive from truthful speech, yielding balanced accuracies near random chance. In contrast, linguistic features, particularly S-BERT embeddings, provide a more discriminative feature space that notably outperforms acoustic baselines. Furthermore, our multimodal experiments reveal that fusing S-BERT embeddings with the selected ComParE features achieves the highest overall balanced accuracy (0.579 using a Linear SVM), whereas combining them with high-dimensional wav2vec embeddings degrades performance. These findings suggest that while deep semantic text features carry the primary predictive weight, carefully selected low-level speech signals harbor complementary paralinguistic cues that successfully enhance deception detection when processed with suitable classification algorithms.
#1321: Context as a Key: Quantifying Verbatim Data Leakage Across Model Scale, Alignment, and Reasoning Architectures
Adam Frémund and Jan Švec
Large Language Models (LLMs) pose severe privacy risks through verbatim memorization, yet the exact impact of prompt context length, safety alignment, and Chain-of-Thought (CoT) architectures remains unquantified in modern open-weight models. We introduce an evaluation framework to quantify data leakage using high-entropy document prefixes (50–250 tokens) to probe state-of-the-art models, including Llama 3.1, Qwen 2.5, Gemma 2, Mistral Nemo, and DeepSeek-R1-Distill. Our analysis yields four key contributions: (1) A Context Multiplier effect, where prefixes longer than 150 token act as activation thresholds that saturate extraction; (2) A Scale-Memorization Correlation paired with an Alignment Tax, demonstrating that while parameter scale deepens memorization, instruction-tuning suppresses shallow leakage; (3) A near-total collapse in verbatim recall for distilled CoT models, which ablation isolates to the distillation phase rather than dynamic inference; and (4) Differential exposure analysis and architectural benchmarking that reveal distinct vulnerabilities, such as elevated data retention risks in Gemma 2. These insights provide critical empirical baselines for developing privacy-preserving conversational AI.
#1326: Contextual Biasing for Streaming Automatic Speech Recognition via Phoneme-Fused Representation and Detect-Then-Inject Decoding
Yun-Ting Sun, Kai-Chen Tsai, Tien-Hong Lo, Jeih-Weih Hung and Berlin Chen
Existing contextual biasing methods for streaming automatic speech recognition (ASR) still face two central challenges: ambiguity between target phrases and phonetically similar distractors, and redundant computation caused by always-on context injection. This work studies a streaming Zipformer-based RNN-T framework with three components: encoder-side contextual biasing, phoneme-fused context representations trained with Guided Attention Cross-Entropy (GA-CE) supervision, and Detect-Then-Inject (DTI) decoding. The phoneme-fused representation with GA-CE yields more robust performance as the distractor set grows, while DTI selectively activates the biasing branch according to a chunk-level margin-based trigger score. Experiments on the LibriSpeech test-other cut-based evaluation show that, under 500 distractors and 320 ms decoding, adding GA-CE on top of the phoneme-fused setting reduces word error rate from 9.89% to 9.73% while also improving U-WER. In the main comparisons, DTI reduces the total real-time factor (RTF) by 31% or more while keeping B-WER broadly competitive and providing a clear accuracy-efficiency trade-off under more challenging conditions.
#1387: Cross-Attention Fusion of Acoustic and Linguistic Features for Parkinson's Disease Detection
Jhon Fredy Mercado-Agudelo, Daniel Escobar-Grisales, Cristian David Ríos-Urrego, Jeferson David Gallo-Aristizábal and Juan Rafael Orozco-Arroyave
Parkinson's disease (PD) causes speech and language disorders, including reduced articulatory precision, monopitch, high semantic variability, and impaired processing of action concepts. This paper presents a multimodal framework that combines window-level Whisper acoustic embeddings with word-level, clinically motivated linguistic features and is trained on PC-GITA and evaluated on an independent test set recorded under uncontrolled conditions. The speech-only model achieves a test AUC of 0.90, while the language-only model achieves 0.63. Cross-attention fusion, in which acoustic window embeddings are enriched with linguistic context, achieves a test AUC of 0.83 with 12 percentage points higher sensitivity than the speech-only baseline, yielding a more balanced sensitivity--specificity profile suited to clinical screening. Attention-weight analysis identifies class-discriminative words present in healthy speech but absent in PD speech, revealing significantly higher motor-semantic and affective activation in healthy controls and linking model behavior to known characteristics of PD-related discourse impairment.
#1404: Cross-Database Performance of Speech-Based Graph Neural Networks for Parkinson’s Disease Detection
Ben Jurkovič, Radim Krupička and Ondřej Klempíř
Speech impairments affect up to 90% of Parkinson’s disease (PD) patients and can serve as early biomarkers of the disorder. Although deep learning approaches for PD speech detection have advanced, their performance might degrade markedly when applied to out‑of‑domain data. This study investigated whether graph structured speech representations can enhance both intra‑dataset and cross‑dataset classification, and evaluated several domain adaptation (DA) strategies for mitigating domain shift. We conducted experiments on seven multilingual corpora covering a variety of speech tasks, including two widely used non-PD dysarthric datasets (TORGO, UASPEECH), employing both conventional acoustic features and pre-trained speech embeddings. The constructed graphs were characterized by homophily ratios and Maximum Mean Discrepancy (MMD). Our analysis showed that most corpora exhibited strong homophily, yet cross‑corpus MMD remained high, indicating substantial distributional divergence that makes transfer learning difficult. Without adaptation, traditional classifiers and graph neural networks achieved near random performance on unseen domains. Applying structural and distributional DA techniques substantially improved results over naive generalization, yielding gains of up to 28% in selected scenarios.
#1323: Cross-lingual Text-to-speech Translation for Regional, Low-resource Languages in Vietnam
Thanh Binh Nguyen, Phuoc Cuong Le, Quoc Hieu Nguyen, Quoc Hung Hoang, Minh Quang Do, Khanh Linh Nguyen and Quang Tuan Nguyen
A major hurdle for tasks involving low-resource languages is the lack of text and speech data. This proves to be even more challenging for cross-lingual text-to-speech translation: converting source-language text into target-language speech. In this paper, we conduct a systematic comparison of representative speech synthesis paradigms for low-resource languages in an example case study of Vietnamese and its regional variant, Muong, by implementing and evaluating three groups of speech models: spectrogram-based, discrete unit-based and hybrid models, under 5.77h of Vietnamese-Muong text-speech data. We then evaluate our experiment for quantitative and qualitative analyses using MCD, LSD, SSIM metrics and a custom MOS test. Our results show that the hybrid architecture achieves the overall best scores at 5.6185 MCD, 0.9268 SSIM, 4.1689 in MOS naturalness and 4.1798 in MOS translation similarity, while the spectrogram-based model shows the best LSD score at 4.8734.
#1333: Data Synthesis and Parameter-Efficient Fine-Tuning for Low-Resource NMT: A Case Study on Q'eqchi' Mayan
Alexander Chulzhanov, Soeren Eberhardt and Arjun Mukherjee
Neural machine translation for digitally low-resource Indigenous languages is often hindered by extreme data scarcity, prompting reliance on extractive web-scraping. To ensure data sovereignty, this study introduces a data synthesis methodology to bootstrap NMT models without scraping target-language parallel text. Focusing on Q'eqchi' Mayan, we transformed community-sourced dictionaries into a massive synthetic corpus, utilizing Parameter-Efficient Fine-Tuning (PEFT) via LoRA adapters on an mT5-base model. In-domain evaluation demonstrates high structural acquisition (BLEU 42.02), proving that synthetic constraints effectively teach complex agglutinative morphology and VOS word order. However, evaluation against an organic glossary reveals a structural-semantic gap (BLEU 0.59), where the model maintains grammatical integrity but lacks the lexical grounding of natural language. The model exhibits overfitting to the constrained structural variance of the synthetic templates; despite high semantic entropy in the pipeline, it struggles with the syntactic fluidity of natural language, forcing organic inputs into rigid learned patterns. Furthermore, an ablation study utilizing a Multi-Task Learning architecture resulted in negative transfer, suggesting that auxiliary tasks competed for limited parameter capacity within the LoRA adapters, causing over-optimization for synthetic markers at the expense of organic flexibility. Ultimately, we establish that synthetic bootstrapping is a highly effective structural primer, but requires authentic data for semantic refinement via Curriculum Learning.
#1361: Early Alzheimer’s Detection Using Siamese Conformer Networks on Slovak Confrontation Naming Tasks
Tomáš Černáček and Milan Rusko
This paper presents an automated analysis of pathological speech utilising confrontation naming tasks extracted from the Slovak ALOIS-DB, a database specifically designed for the early detection of Alzheimer’s disease. Confrontation naming exercises pose a significant computational challenge, as vocalisations from healthy controls and patients with mild cognitive impairment are often acoustically indistinguishable and show no overt clinical differences in standard speech features. To capture subtle, latent pathological markers in these short-duration signals, we propose a Siamese Neural Network architecture employing a Conformer encoder to process high-resolution spectrograms. Our approach integrates Feature-wise Linear Modulation (FiLM) to dynamically condition representations based on specific naming subtasks. By optimising a contrastive loss function, the model learns a robust metric space in which speech samples are clustered by their clinical characteristics rather than linguistic content. Experimental results demonstrate that Siamese-based architectures can distinguish pathological patterns in isolated words, offering a viable tool for non-invasive, objective screening in the Slovak language context.
#1381: Evaluating the Effect of Fine-Tuning on Usability Annotation of English-Slovak Translation
Martin Chotár, Matúš Kleštinec and Daša Munková
This study investigates the viability of using large language models (LLMs) as automated evaluators of machine translation usability in the English--Slovak language pair. Focusing on a literary-biographical corpus about Seamus Heaney and its Slovak machine translation, the study examines whether LLM-based evaluators can approximate the judgments of professional human translators when assessing the pragmatic usability of translated segments. A 100-segment Golden Standard was independently annotated by professional translators and compared with evaluations produced by multiple LLMs, including base, fine-tuned, heuristic, and reasoning-based variants. The analysis combines repeated-measures ANOVA, post-hoc homogeneous group classification, and rank-based correlation measures to assess differences in scoring severity and evaluator alignment. The results show that human annotators apply a stricter usability standard than most off the shelf LLMs, which tend to exhibit a benevolence bias by overestimating translation quality. Domain specific fine-tuning substantially reduces this bias, aligning selected models with the human evaluation cluster. The findings also indicate that explicit reasoning prompts improve the strictness of smaller models, whereas frontier models appear capable of approximating human judgments without additional deliberation. Overall, the study demonstrates that properly calibrated LLM evaluators can provide a reliable and scalable alternative or complement to human annotation in usability oriented assessment of machine translation for morphologically rich languages.
#1348: Evaluating the Sufficiency of Verbal Picture Description Task for Cognitive Impairment Detection
Tomáš Lebeda, Jan Švec and Luboš Šmídl
Early detection of cognitive impairment is essential for timely intervention, yet clinical assessments remain time-consuming and expensive. Recent automated approaches based on speech and language analysis offer a promising alternative, but they often rely on a relatively large number of cognitive tasks, which can increase examination time and patient burden. In this work, we investigate whether reliable cognitive screening can be achieved using only two complementary tasks: complex scene description and subsequent object recall. Both tasks are derived from the DigiDiaDem dataset and jointly target key cognitive domains, including language production, executive function, and short-term memory. Spoken responses are processed using an automatic speech recognition system followed by a custom semantic parsing pipeline. From the resulting representations, we extract interpretable features such as semantic loss vectors, description richness measures, and lexical statistics. The problem is formulated as a binary classification task distinguishing cognitively healthy and impaired participants. We evaluate the proposed approach using logistic regression as the primary classifier, complemented by additional models, within a framework that includes feature selection, hyperparameter optimization via successive halving, and nested cross-validation. The performance of individual tasks and their combination is compared against a baseline derived from the full task set. The results demonstrate that the combination of the two selected tasks achieves strong classification performance, suggesting that a substantially simplified test battery may still provide sufficient diagnostic information. These findings support the feasibility of shorter, more efficient cognitive screening protocols based on speech analysis.
#1356: Evaluation of Transformer Language Models for Hate Speech Detection in Croatian Online Text
Duje Jurić, Teo Matošević, Teo Radolović and Marina Bagić Babac
Hate speech detection for Croatian remains underexplored, although Croatian represents a linguistically meaningful case for abusive-language research as a low-resource and morphologically rich language. The task is further complicated by the frequent use of indirect, coded, or context-dependent expressions of abuse, which limits the usefulness of simple surface cues. In this paper, we compare sparse lexical baselines and several transformer models on the FRENK dataset of moderated Croatian news comments. The evaluated models include multilingual encoders as well as BERTić, which is pretrained on closely related South Slavic languages. The results show that BERTić achieves the strongest performance, reaching 0.811 ± 0.006 F1-macro, and remains robust across random seeds while outperforming both sparse baselines and multilingual transformer alternatives. Additional analyses indicate that the dataset is dominated by relatively short comments and that longer input windows provide little benefit. The findings support the importance of related-language pretraining for Croatian hate speech detection and suggest that simple lexical augmentation cannot substitute for contextual modeling.
#1327: From Paper to Speech: ASR-Driven Digitalization of the PICNIR Cognitive Test
Filip Polák, Jan Tupý, Luboš Šmídl and Aleš Bartoš
Early detection of memory impairment requires sensitive, rapid, and scalable assessment tools. One such instrument is the Picture Naming and Immediate Recall test (PICNIR), a brief cognitive screening method based on naming 20 black-and-white drawings followed by immediate recall. Its in-person administration takes 3–4 minutes and is validated for very mild cognitive impairment. The digitalization of PICNIR offers the potential for remote, scalable, and automated cognitive screening. In this study, we introduce an electronic version of PICNIR and evaluate whether its performance is comparable to the traditional in-person format. In contrast to the paper-based version, the electronic PICNIR presents stimuli sequentially, requiring participants to name each drawing aloud and subsequently recall as many items as possible. We describe the system architecture of the DigiDiaDem application, including a custom-designed Automatic Speech Recognition (ASR) module tailored to the target population. Beyond replicating standard clinical scoring, the transition to a speech-based digital format enables the extraction of novel acoustic, temporal, and linguistic features—such as response latency, hesitations, and pause patterns—which represent sensitive markers of cognitive decline inaccessible in traditional assessments. Our findings suggest that the ASR-driven electronic PICNIR provides a reliable, scalable, and diagnostically enriched alternative to conventional cognitive screening methods.
#1319: GRASS - HRI: A Corpus of Spontaneous Austrian German Dialogues with the Social Robot Furhat
Sophie Lennkh, Michael Paierl and Barbara Schuppler
This paper presents the first corpus of spontaneous robot-directed dialogues for Austrian German. For the data collection, 20 native speakers of Austrian German (10 female, 10 male, 0 non-binary) conducted five conversations each with the social robot Furhat playing "Who am I?", resulting in approximately 11 hours of speech and 2386 utterances. Recordings are available in high audio quality from the head-mounted microphone and in addition as recorded by Furhat itself. Recordings were manually orthographically transcribed and time-aligned, maintaining the original temporal structure of the dialogues. During the experiment, five different state-of-the-art automatic speech recognition (ASR) systems were used, resulting in differences in quality of the system responses and in user satisfaction. Also the ratings of the users experience are available together with the corpus. Our motivation for making the collected data available actually comes from our observations from our detailed analysis of Word Error Rates for the different ASR systems, described in the second part of the paper: All systems, even those fine-tuned on Austrian German, performed significantly worse than on spontaneous human-human conversational speech corpora, and none of them fulfilled the necessary real-time requirements.
#1324: Gender Bias in Greek Pronoun Resolution: Evaluating Multilingual and Language-Specific LLMs
Konstantinos Diamantopoulos and Stergios Chatzikyriakidis
We present the first systematic study of gender bias in Greek pronoun resolution using large language models. We introduce a novel dataset of 906 sentence pairs exploiting Greek's morphological property that certain epicene nouns encode grammatical gender solely through articles---which we deliberately omit to create genuine ambiguity. Evaluating four LLMs including the Greek-specific Llama-Krikri with six prompting strategies, we find a surprising performance-fairness trade-off: Krikri achieves the highest ambiguity recognition (79.2%) yet exhibits the largest gender-based inconsistency (26.1%, Cohen's d=0.68). Prompting strategy variation (78.9 percentage points) exceeds model variation (62.8 points), with few-shot prompting achieving 68.3% ambiguity recognition while maintaining only 5.0% gender bias---demonstrating that the performance-fairness trade-off can be substantially mitigated through strategic prompt engineering.
#1320: Is this Funny to You?: A Reader-Aware Humour Classification Dataset
Kennedy Roland and Milton King
Humans tend to encounter humour in different ways, creating many individual, unique experiences. Jokes are created to be humorous but are dependent on the individual's preference, as things aren't considered to be globally funny or not funny. This research aims to determine if we can predict if someone will find a particular joke humorous based on personal characteristics. We propose the Is this Funny to You (IFY) dataset, which is a novel dataset that contains reader-specific annotations of jokes and personal characteristics about the readers. We evaluate personalized reader-aware classifiers to determine if a specific individual will find a joke funny based on some of their demographic information and other joke preferences. We test these models on different scenarios, which differ on having seen the same reader and jokes during the training and testing or not.
#1370: LLMs as Linguistic Experts: The Case of Morphological Segmentation
Vojtěch John and Zdeněk Žabokrtský
Morphological segmentation is a long-standing linguistic task. Due to data sparsity, its difficulty is especially pronounced in low-resource settings, which has recently motivated work on using large language models (LLMs) for morphological segmentation in zero-shot or few-shot scenarios. However, prior studies are usually limited in scope, focusing on a single model and few languages, thus leaving comparative performance of LLMs and its underlying factors insufficiently understood. In this paper, we evaluate 5 state-of-the-art LLMs on a small manually morphologically segmented benchmark in 13 languages. We perform several experiments in both a zero-shot and a few-shot setting. The models achieved overall morph boundary F1 between 28% (LLaMa 3.1) and 78% (Claude 4.5 Sonnet). Through detailed error analysis, we identify several key factors influencing model performance. The results suggest model performance strongly varies across languages, and that although the few-shot setting leads to increased performance, the models make limited use of additional linguistic information.
#1359: Large Language Models for Norwegian Bokmål-Nynorsk Translation: Scaling Laws and the Limits of Back-Translation
Lars Bungum, Oscar Andreas Halvorsen and Jon Atle Gulla
This paper explores machine translation between the two written forms of Norwegian, Bokmål and Nynorsk. While rule-based systems have long been state-of-the-art for this language pair, we show that fine-tuned LLMs have now overtaken them. Norwegian-specific models consistently outperform both the Apertium baseline and general-purpose multilingual LLMs across SacreBLEU, BERTScore, and COMET. Notably, cold-start Norwegian models perform nearly as well as their warm-start counterparts. Model size is a strong predictor of translation quality, explaining 68--79% of performance differences. Back-translation, however, yields no improvement and leads to a modest decline in performance. Although our best LLM still comes up short of the dedicated encoder-decoder model NorT5 on the NorBench benchmark, the results demonstrate that decoder-only LLMs are now viable for close language pairs.
#1351: Large Language Models for Sparse Entity Alignment over Knowledge Graphs
Péter Kardos, Máté Vass, Miklós Krész, and Richárd Farkas
Entity Alignment (EA) identifies entities across Knowledge Graphs (KGs) that refer to the same real-world object. While existing benchmarks assume dense alignment with near-complete overlap, real-world scenarios often exhibit sparse overlap where only a small fraction of entities have matching pairs. We address sparse entity alignment using a Retrieval-Augmented Generation (RAG) pipeline: embedding-based candidate retrieval followed by selection using embedding reranking or Large Language Models (LLMs). We introduce textual entity representation strategies combining node attributes with graph neighborhood information. We evaluate our approach on the OAEI Knowledge Graph Track, providing the first systematic evaluation of LLM-based methods for sparse entity alignment. Results show that representation design strongly influences alignment performance. LLM-based selection achieves highest accuracy by effectively rejecting non-matching candidates, while embedding-based methods remain competitive with richer representations.
#1367: Large-Scale Computational Morphology and Lexical Resources for {L}atvian and {L}atgalian
Lauma Pretkalniņa, Krasimir Angelov, Kristīne Pokratniece and Normunds Grūzītis
In this paper, we tackle the morphological analysis and synthesis, and interlingual lexical mapping of two morphologically rich Baltic languages -- Latvian (1.5M native speakers) and Latgalian (indigenous, 150k native speakers, under-resourced). We believe that, in the era of large language models, rule-based models are crucial for countering data sparsity problems arising from a large diversity of morphological forms and comparatively small amount of text corpora available for such languages, especially when it comes to controlled language generation (NLG). We describe the leading open-source morphology model for Latvian and our work adapting this model to Latgalian. We also describe the integration of the morphology models with the largest multi-functional lexical database for Latvian and Latgalian -- Tezaurs.lv. We have developed a method for automated porting of inflectional paradigms and lexical entries from Tezaurs.lv to Grammatical Framework (GF) -- a computational infrastructure and a resource grammar library for controlled multilingual NLG. The lexical mapping among languages is enabled via linking to WordNet, which is supported by both Tezaurs.lv and GF. Together, these fundamental computational language resources allow for rapid development of knowledge-based lexical and grammatical applications for low-resource languages, especially when predictable NLG is crucial.
#1379: Leveraging Zero-Shot TTS for Data Augmentation: A Comparison of VITS and StyleTTS2 in Low-Resource and Low-Quality Conditions
Zdeněk Hanzlíček, Jindřich Matoušek, Daniel Tihelka and Jan Lehečka
The performance of many Text-to-Speech systems still depends on the quality and volume of training data. In many practical situations, the appropriate data is not available. This paper describes our experiments on using the zero-shot capabilities of the ZipVoice TTS system to create a synthetic training corpus for other TTS architectures, namely VITS and StyleTTS2. A comparative analysis revealed that 1) ZipVoice can generate cleaner speech than the default noisy prompt, and 2) both VITS and StyleTTS2 were preferred to ZipVoice in listening tests. These findings suggest that generative data augmentation is a viable approach for developing TTS systems when appropriate human data is unavailable. This paper is a case study that aims to develop a stable personal TTS system for a prominent Czech academic artist, utilizing his low-quality archival recordings.
#1335: LibriConvo: Simulating Conversations from Read Literature for ASR and Diarization
Máté Gedeon and Péter Mihajlik
We introduce LibriConvo, a synthetic conversational speech corpus for speaker diarization and automatic speech recognition (ASR), built by instantiating the previously proposed Speaker-Aware Simulated Conversation (SASC) framework in a dataset and benchmarking setting. The main contribution of this paper is a corpus construction pipeline and benchmark derived from that framework. To make the data more suitable for downstream ASR and diarization, conversational timing statistics are estimated from English CallHome using external voice activity detection, long pauses are compressed, LibriTTS utterances are grouped by book to improve local semantic continuity, and room impulse responses are selected with a spatial-plausibility heuristic. The resulting corpus contains 240.1 hours of audio across 1,496 dialogues involving 830 speakers, partitioned into speaker-disjoint train, validation, and test splits. We report baseline results for both diarization and ASR. On the test split, Sortformer outperforms the pyannote pipeline in diarization (11.1% vs. 24.4% DER). For ASR, a Fast Conformer-CTC XLarge model fine-tuned with Serialized Output Training achieves 7.29% WER and 6.97% cpWER, outperforming zero-shot Whisper-large-v3. These results position LibriConvo as a practical benchmark for studying synthetic conversational speech and for evaluating multi-speaker speech processing systems.
#1311: Location-Aware Language Models via Secondary Embeddings
Gokul Srinivasagan and Munir Georges
Pretrained transformer-based language models achieve strong performance across a wide range of NLP tasks but remain limited in encoding geo-locational semantics, leading to suboptimal representations of place names and spatial entities. In this work, we propose a lightweight, model-agnostic approach for injecting geo-spatial awareness into pretrained embeddings without modifying the tokenizer or requiring costly retraining. Our method augments input representations with structured geographic signals by combining location names with their corresponding latitude and longitude, and employs a location-focused masking to better align textual representations with real-world spatial relationships. This design allows the model to incorporate geo-spatial context while preserving existing semantic and syntactic knowledge. Experimental results demonstrate substantial improvements in geo-spatial alignment while maintaining comparable performance on standard NLP benchmarks such as GLUE. The method is computationally efficient, requiring only minutes of additional training, and generalizes across multiple model architectures and scales.
#1357: Modelling Thematic Structure in Clinical Dialogues for Interpretable Depression Detection
Esaú Villatoro-Tello, Ernesto Antonio Reyes-Ramírez, Sergio Burdisso, Hasindri Watawana, Petr Motlíček, Fernando Sánchez-Vega and Adrian Pastor López-Monroy
Detecting depression from clinical interviews requires models that capture discourse-level thematic structure rather than relying solely on lexical features. We propose Topic-GCN, a heterogeneous graph convolutional network that jointly encodes word, latent topic, and document nodes, enabling explicit modelling of topic--document associations for interview-level classification. The learned topic nodes surface interpretable symptom-relevant clusters whose class-conditional relevance can be mapped to established psycholinguistic categories. Experiments on the DAIC-WOZ and E-DAIC corpora show that Topic-GCN improves predictive performance over state-of-the-art methods while preserving interpretability. A post-hoc analysis using LIWC categories shows that the learned topic structure aligns with relevant symptom dimensions. (Code: https://github.com/idiap/Topic-GCN_for_depression_detection)
#1345: Multimodal Speech Recognition in High-Noise Factory Floors for Human Robot Collaboration
Juan Camilo Vásquez-Correa, Eneko Tomé, Haritz Arzelus, Ana Díaz de Zugazúa, Sara Sillaurren and Aitor Álvarez
Natural language interaction in industrial settings is severely hindered by extreme acoustic noise, which makes standard Automatic Speech Recognition (ASR) systems unreliable. To enable flexible human-robot collaboration, there is a critical need for systems able to interpret operator commands under harsh conditions without requiring expert programming. This paper proposes a novel architecture that integrates an ASR module in parallel with a Visual Speech Recognition model to enhance transcription accuracy under adverse acoustic conditions. To resolve potential discrepancies between the two streams, a Large Language Model (LLM)-based post-processing module acts as a semantic arbitrator, evaluating the most probable transcription based on contextual and industrial task constraints. The framework was tested in an industrial scenario characterized by high non-stationary noise. The multimodal approach, combined with LLM-based arbitration, outperformed both unimodal ASR and VSR baselines across various scenarios. By combining auditory and visual cues with the reasoning capabilities of LLMs, this work provides a solution for hands-free robot programming in noisy manufacturing plants, bridging the gap between human intent and robotic execution.
#1328: OLaPh: Optimal Language Phonemizer
Johannes Wirth
Phonemization is a critical component in text-to-speech synthesis. Traditional approaches rely on deterministic transformations and lexica, while neural methods offer potential for higher generalization on out-of-vocabulary (OOV) terms. We introduce OLaPh (Optimal Language Phonemizer), a hybrid framework that integrates extensive multilingual lexica with advanced NLP techniques and a statistical subword segmentation function. Evaluations on the WikiPron benchmark show OLaPh significantly outperforms established baselines in overall accuracy and maintains robustness on OOV data through advanced fallback mechanisms. To further explore neural generalization, we utilize the framework to synthesize a high-consistency training corpus for an instruction-tuned Large Language Model (LLM). While the deterministic framework remains more accurate overall, the LLM demonstrates strong generalization, matching or partly exceeding the framework's performance. This suggests that the LLM successfully internalized phonetic intuitions from the synthetic data that transcend the framework's capabilities. Together, these tools provide a comprehensive, open-source resource for multilingual grapheme-to-phoneme conversion (G2P) research.
#1386: On the Explainability of Speech Impairment Regarding Dysarthria Progression in Parkinson's Disease
Jeferson David Gallo-Aristizábal, Cristian David Ríos-Urrego, Daniel Escobar-Grisales, Jhon Fredy Mercado-Agudelo and Juan Rafael Orozco-Arroyave
Parkinson's disease (PD) commonly affects speech motor control and is often associated with hypokinetic dysarthria, characterized by impairments in articulation, phonation, and prosody, which may reduce speech intelligibility. Most studies on speech analysis in PD have primarily focused on differentiating between patients with PD and healthy control (HC) subjects. Additionally, there has been limited research on evaluating disease severity while also considering the interpretability of linguistic information. Therefore, this work explores whether self-supervised speech representations can support the automatic assessment of dysarthria severity using the modified Frenchay Dysarthria Assessment (mFDA). wav2vec2.0 embeddings were extracted from read-text recordings and grouped at different temporal and linguistic granularities, including frame, phoneme, phoneme-transition, syllable, word, sentence, and subject levels, using forced alignment. Results balanced accuracy values between 50.4% and 54.1%, exceeding the chance level of 25%. In addition, the analysis of specific linguistic units and phonological classes supports the idea that severity-related information is concentrated in articulatory patterns involving plosives, fricatives, nasals, and consonant-vowel transitions, where fine control of articulators such as lips, tongue, and velum is required.
#1353: On the Influence of VITS Initialisation on Speech Quality
Daniel Tihelka and Zdeněk Hanzlíček
The paper describes the continuation of a saga on experiments with the VITS TTS model. Now we focus on the quality of speech generated by models fine-tuned from foundation models, which were trained from data with various styles and sizes. Four different speech corpora were used to build the foundation models, the amateur recordings, audiobooks read by actors, podcasts with natural and causal speech style, and the combination of all these. Although the results are not clearly conclusive, it seems that a v{DNA} of the foundation corpora may influence the final model.
#1358: On the Robustness to Recording Condition of Speaker and Accent Embeddings for a Voice Cloning System
Philippe Martin, David Guennec, Vincent Barreaud, Nelly Barbot and Damien Lolive
Speech attribute embeddings for zero-shot Text-To-Speech and Voice Conversion must ensure robustness to acoustic perturbations and strict attribute disentanglement. However, current models frequently encode undesirable channel artifacts. This paper introduces an evaluation protocol to systematically quantify the resilience and disentanglement capabilities of embeddings to such perturbations. The protocol is applied to four popular feature extraction models (speaker: Spk ECAPA-TDNN, X-Vector; accent: Acc ECAPA-TDNN, GenAID) using audio from LibriTTS and SAAD corpora, which we subject to controlled acoustic degradations. Analyzing latent space integrity via classification F1-scores, clustering, and dispersion metrics reveals disparities between models. For speaker embeddings, while Spk ECAPA-TDNN demonstrates high resilience to environmental perturbations, X-Vector's latent space collapses, encoding environmental signatures instead of identity. While Acc ECAPA-TDNN and GenAID accent models exhibit a vulnerability to environment, GenAID appears to be more sensitive, especially regarding severe background noise. Furthermore, Acc ECAPA-TDNN fails to dissociate speaker identity from accent features, whereas GenAID successfully neutralizes speaker-specific information, highlighting the critical need for more robust accent representations in degraded conditions. Consequently, GenAID relies less on accent information and provides a more homogeneous performance when subjected to environmental noise.
#1350: On the Synthesis of Dysarthric Speech: Evaluating Zero-Shot and Fine-Tuning Approaches
Jindřich Matoušek, Luboš Šmídl, Paula Andrea Pérez-Toro, Elmar Nöth, Jan Lehečka, Andreas Maier and Tomás Arias-Vergara
The critical scarcity of pathological speech data significantly hinders the development of robust automatic speech recognition (ASR) systems for individuals with speech impairments. To address this issue, this paper investigates the capabilities of modern neural text-to-speech (TTS) architectures to synthesize highly accurate Spanish dysarthric speech. We explore two distinct synthesis strategies (zero-shot voice cloning and model fine-tuning) and two TTS models (ZipVoice and StyleTTS2) applied to recordings of 50 speakers with Parkinson's Disease and 50 healthy controls. To evaluate the quality and preservation of speaker identity in the generated samples, we employ WavLM-based speaker embeddings alongside detailed acoustic-phonetic analyses. By systematically comparing real and synthesized speech, we aim to determine the optimal TTS training approaches for capturing dysarthric characteristics, thereby establishing a crucial foundation for future ASR data augmentation efforts. Experimental results demonstrate that model fine-tuning significantly outperforms zero-shot approaches, with the fine-tuned StyleTTS2 emerging as the most effective configuration, achieving the highest speaker similarity and demonstrating superior robustness across classification tasks and phonetic/prosodic analyses.
#1410: OnkoRAG: Modelling Cross-Lingual German-Arabic Medical Question Answering
Nicholas Gareth Jennings, Volha Petukhova and Dietrich Klakow
Language barriers, domain-specific complexity, and limited access to reliable information hinder informed healthcare decision-making. To address these challenges, this paper presents a corpus, an LLM-based framework, and an evaluation methodology for cross-lingual German-Arabic medical question answering (QA). We introduce a German-Arabic oncology QA benchmark that models the linguistic behavior of native speakers and Arabic-speaking learners of German through synthetic error simulation. Along with constructing a Standard Arabic oncology QA dataset, we generate learner variants using back-translation and noise injection to emulate characteristic non-native errors. To mitigate LLM hallucinations, i.e., plausible yet factually inaccurate responses, we adopt retrieval-augmented generation (RAG) in a cross-lingual setting. The resulting OnkoRAG system is evaluated in terms of retrieval performance, relevance, faithfulness, and factual correctness. The results highlight the challenges of cross-lingual RAG, with recall dropping from 0.80 in German to 0.68 in Arabic. Injected noise has only a minor impact, reducing ROUGE scores by 0.02. Although faithfulness remains high for both languages (0.83), factual correctness is comparatively low, reaching 0.51 and 0.48 for German and Arabic, respectively. These findings underscore the difficulty of ensuring factually accurate responses in cross-lingual medical QA despite strong retrieval and highly faithful generation.
#1412: Online Punctuation and Capitalization Restoration Using Seq2Seq Approach
Martin Poláček, Petr Červa and Lukáš Matějů
This paper presents a lightweight Seq2Seq approach to automatic punctuation and capitalization restoration (APCR) for real-time streaming applications. Our method employs a reduced BART-Small models (for Czech and English) that processes input with minimal right context of only five words, enabling low-latency processing suitable for CPU deployment. It simultaneously restores punctuation (periods, commas, question marks) and capitalization through iterative decoding, allowing flexible character modifications within words, e.g., us versus U.S. We evaluate it across various datasets, comparing it with our previous ELECTRA-based classification system and NeMo framework. Our results demonstrate significant improvements. For example, in the automatic punctuation restoration (APR) task on automatic speech recognition (ASR) outputs, the weighted-average F1 score reaches 77.7%, corresponding to a 6.3 percentage-point improvement over the ELECTRA baseline.
#1352: Optimizing Streaming Zipformer for Czech Humanoid Robotics: Balancing Accuracy, Latency, and Ego-Noise
Jan Lehečka, Josef V. Psutka, Aleš Pražák and Josef Psutka
Natural human-robot interactions require Automatic Speech Recognition (ASR) systems that are not only accurate but also highly responsive. However, streaming models often exhibit "lazy emission" behavior, where latency increases as the model converges toward higher accuracy. This paper presents an optimized streaming ASR system for the Czech language based on the Zipformer architecture, specifically tailored for humanoid robotic platforms. We investigate the trade-off between recognition performance and latency by employing a calibrated delay penalty during training. Our results demonstrate that a penalty of 0.001 effectively regularizes emission latency with negligible impact on the ASR errors. Furthermore, we evaluate the system’s deployment on a limited edge device (NVIDIA Jetson Xavier AGX), showing that not only does the system run smoothly on less than half of a single CPU core, but it also provides superior responsiveness over GPU-based inference. Finally, we analyze the impact of mechanical ego-noise across different microphone placements, identifying the proximity of head motors to the microphone as a primary source of performance degradation in dynamic robotic states. Our final configuration achieves a WER of 6.4% and a total perceived latency below 400 ms, providing a robust benchmark for real-time speech processing in Czech mobile robotic systems.
#1391: Pseudonymisation for Morphologically Rich Languages
Rinalds Vīksna and Mārcis Pinnis
Anonymisation or pseudonymisation of personal information (PI) is a core requirement in privacy-preserving natural language processing applications. Existing approaches typically rely on masking techniques or pseudonymise PI using replacement lists or random surrogates. For analytic languages, such as English, this approach works well; however, in morphologically rich languages, where words exhibit complex inflectional paradigms, pseudonyms must agree in case, number, and gender with the syntactic role in the sentence. We introduce a method that produces pseudonyms for PI entities while maintaining grammatical compatibility with the surrounding text. We evaluate our approach on named entity recognition and sentiment analysis tasks in English and Polish. Our results show that the data pseudonymised using our morphology-aware pseudonymisation method remains useful for the downstream tasks.
#1364: Reading between the Lines: Leveraging Large Language Models for Global Dementia and Depression Assessment from Clinical Interviews
Franziska Braun, Alea Rüggeberg, Thomas Ranzenberger, Hartmut Lehfeld, Thomas Hillemacher, Tobias Bocklet and Korbinian Riedhammer
Dementia and depression are the most prevalent neuropsychiatric disorders in geriatric populations, and their overlapping symptoms pose major challenges for differential diagnosis. In this study, we investigate open-weights Large Language Models (LLMs) for predicting dementia and depression severity from speech samples collected during standardized history taking interviews with 154 German-speaking subjects. We introduce an observer-based Global Depression Scale (GDS-D) aligned with the established Global Deterioration Scale (GDS), enabling parallel global staging of affective and cognitive symptoms. We compare three LLMs (Mistral 3.1, DeepHermes, Qwen3) in two settings: (1) zero-shot prediction and (2) LLM-based feature extraction for Support Vector Regression, using human and pause-enriched transcripts. Results show that LLMs effectively predict depression severity in zero-shot settings (best MAE of 0.60), while dementia assessment benefits substantially from structured feature extraction (best MAE of 0.78), reducing errors by up to 35% over zero-shot baselines. Pause-enriched transcripts achieve competitive performance with human transcriptions, demonstrating the viability of fully automatic screening pipelines for differential neuropsychiatric assessment.
#1322: SYNODA: Synthetic Data-Oriented Domain Adaptation for Automatic Speech Recognition
Lukas Graf and Christian Bergler
Automatic Speech Recognition for healthcare remains constrained by the limited availability of medical acoustic data. Clinical speech data is highly sensitive, often contains personal health information, and is subject to strict privacy regulations, making real-world medical audio difficult to collect, share, and use at scale. To address this challenge, we introduce SynGMA, a synthetic medical audio dataset, prototypically implemented for German, as a proof of concept for a language-independent approach to domain-specific speech recognition in healthcare. SynGMA is generated through the combination of large language model-based text generation followed by downstream Text-To-Speech synthesis, enabling the scalable creation of realistic, privacy-preserving medical recordings. In total, SynGMA comprises 34,062 audio snippets with 57 hours of synthetic medical speech. To assess its practical value quantitatively, an Automatic Speech Recognition system is fine-tuned on SynGMA and evaluated against a non-adapted baseline. The results show that domain adaptation with synthetic medical audio leads to a 19.82% absolute reduction in Word Error Rate, demonstrating substantial gains in recognizing domain-specific utterances. At the same time, performance on general language-specific speech remains largely unaffected, indicating that fine-tuning for the medical domain does not come at the expense of broader usability. These findings demonstrate that synthetic data generation combined with downstream domain adaptation is a promising and transferable approach for extending Automatic Speech Recognition to sensitive or data-scarce domains beyond healthcare.
#1316: SYN_r25: a Corpus Linked to a Database of Multi-Word Expressions
Tomáš Jelínek and Hana Skoumalová
This paper presents the SYN_r25 corpus, a large corpus covering all six 100-million-word representative corpora of written Czech published by the Czech National Corpus since 2000. This corpus has been enriched with detailed annotation of multiword expressions (MWEs). For the first time in the Czech corpus linguistics, MWEs are systematically linked to a structured lexicon—the LEMUR database—which provides comprehensive linguistic and computational descriptions of phraseological units. We review previous approaches to MWE annotation, including the FRANTA system, and describe the design of LEMUR, which captures syntactic, semantic, and idiomatic properties of MWEs. In SYN_r25, each MWE is annotated with simplified attributes such as lemma, syntactic type, transformation, and idiomaticity dimensions. These annotations are accessible via the KonText interface, where each MWE is linked to its full entry in the LEMUR database, allowing for both linguistic analysis and natural language processing applications. The corpus will remain stable after publication, while the lexicon will continue to evolve. SYN_r25 marks a significant advancement in Czech corpus annotation and provides a robust foundation for research in phraseology and computational linguistics.
#1396: Size Matters: Foundation Model for Czech HTML Documents
Martin Dvořák, Vít Tlustoš, Artyom Voronin, Martin Habrovec, Kateřina Podlesná, Barbora Rišová, and Josef Vonášek
Creating universal, high-quality representations of web documents in high-traffic industrial environments requires models that are both performant and economic. Existing approaches, however, often depend on large models, overlook the structural information inherent in HTML, or are constrained by short context windows, limiting their ability to process real-world web pages. We present HTML-LM, a compact foundation model with 154 million parameters that addresses these limitations through HTML-aware training and a ModernBERT-based architecture. It was trained on 100 million web documents using multiple objectives, including masked language modeling, bag-of-words prediction, and contrastive distillation from large language models. Consequently, HTML-LM sets a new state-of-the-art for classification and regression applications in the Czech Internet domain, surpassing both larger encoders and small-sized LLMs. The model is deployed in production, processing thousands of web documents per second, and released to the community under the CC BY-NC 4.0 license. https://huggingface.co/Seznam/html-lm
#1403: SofiaFala ECOA: A Framework for Multimodal Corpora Construction for Speech Disorder Analysis and Assistive AI Systems
Igor Lourenço Bartijotto Manzano Bento, Stephany Xesquevixos, Yingqiang Gao, Patricia Puppin Mandrá and Alessandra Alaniz Macedo
Speech is a fundamental medium for human communication and expression, yet individuals with speech disorders often face significant barriers to social interaction, education, and autonomy. Despite the current advances in artificial intelligence (AI), the scarcity of diverse, high-quality atypical speech corpora in Brazilian Portuguese still limits the development of inclusive and robust assistive tools. This paper introduces SofiaFala ECOA, a framework designed to build an unprecedented multimodal corpus of atypical speech grounded in the principles of digital sovereignty and citizen science. SofiaFala utilizes a responsive platform for participatory collection of audio and additionally also video data, and an administrative dashboard for expert-driven classification and pre-processing of the collected material. Following the FAIR principles (Findable, Accessible, Interoperable, and Reusable), our framework involves a standardized protocol where participants record words and phrases covering all phonemes of the Brazilian Portuguese language, accompanied by anonymized clinical and demographic metadata. After one year, we collected 2,261 processed audios from 165 donors with predominance of children (average age of 8 years), from which 1,372 samples were classified as Good Quality by speech therapy specialists, while other categories such as Noise, Delayed Start, and Non-verbal etc. productions were also identified and mapped. This initiative not only provides essential resources for AI research in clinical settings such as automatic speech recognition (ASR), but also empowers individuals with communication disorders by promoting social inclusion and self-determination.
#1344: Syllable Stress Detection to Evaluate Pathological Speech and L2 Language Level Pronunciation
Juan Camilo Vásquez-Correa, Haritz Arzelus and Aitor Álvarez
Syllable stress is a major key aspect of speech frequently altered by neurological impairments and varying levels of non-native language proficiency. This paper presents a novel syllable stress detection pipeline designed to evaluate pathological speech and assess L2 learner pronunciation. Separate models for Spanish and English were trained using Common Voice and Librispeech corpora. The proposed architecture extracts multidimensional syllable-level acoustic representations, combining prosodic, spectral, and articulatory features. These representations are processed through a 1D-convolutional projection and a Transformer encoder to capture inter-syllabic dependencies, followed by a time-distributed multi-layer perceptron for per-syllable stress prediction. In-domain evaluations demonstrated overall accuracies of 97.5% for Spanish and 93.9% for English. Furthermore, the model's practical utility was validated through benchmark evaluations on out-of-domain datasets, spanning L2 Spanish learners and clinical populations with Aphasia, Alzheimer’s, and Parkinson’s diseases. The model successfully captured deviations in stress patterns, with detection accuracy reflecting both Aphasia severity and Spanish L2 proficiency levels. These findings suggest that automated syllable stress detection serves as a viable, objective tool for clinical speech assessment and language learning tracking.
#1362: Tremor Vision: Cross-Domain Multiclass Plosive Burst Detection from Audio-Derived Spectrogram Images for Parkinsonian DDK Speech Analysis
Sajjad Hadi, Christopher Hauer, Philipp Klumpp, Elmar Nöth, Andreas Maier and Tomás Arias-Vergara
Diadochokinesis (DDK) is a rapid syllable repetition task (for example, PA-TA-KA) widely used as a clinical marker of motor speech impairment in Parkinson's disease. Most diadochokinesis pipelines depend on hand-crafted energy thresholds or template matching, which are brittle under noise and pathological speech distortions. We present Tremor Vision, a fully data-driven pipeline that reformulates plosive burst localization as an object detection task on three-channel audio-derived image patches formed from a mel spectrogram, spectral flux, and waveform envelope. YOLO26x is pre-trained on an age-balanced subset of the multilingual CommonPhone corpus and fine-tuned on a Spanish clinical corpus of healthy controls and Parkinson's disease patients. On the unseen Parkinson's disease group, the model achieves mAP@0.50 = 0.433, a 57.5% relative improvement over training from scratch (0.275). For the clinically relevant classes /k/ (KA), /p/ (PA), and /t/ (TA), mAP@0.50 reaches 0.786, 0.571, and 0.812 respectively. The pipeline requires no manual signal processing and delivers per-subject burst-level bounding boxes. Furthermore, the model demonstrates sensitivity to articulatory imprecision by correctly localizing non-canonical substitution tokens (e.g., PA-BA) mapped to their actual acoustic classes, establishing a foundation for downstream DDK feature extraction and dysarthria severity assessment.
#1342: Turn-by-Turn Acoustic-Prosodic Alignment and its Relationship to Lexical and Semantic Similarity in Casual Conversations
Jana Winkler and Barbara Schuppler
One prevalent characteristic of conversational speech is the alignment between speaker pairs. Whereas previous works have predominately analyzed lexical and acoustic dimensions of alignment separately, this work focuses on analyzing their interaction on a turn-by-turn basis. Based on 851 turn pairs extracted from a corpus of casual Austrian German conversations and using Conditional Inference Trees for modeling, we find that whereas higher lexical similarity scores significantly correlate with pronunciation matching, articulation-rate alignment and the absence of overlapping speech, higher semantic similarity scores tend to co-occur significantly with pitch-minimum alignment and, for longer IPUs, articulation-rate alignment. Our quantitative findings on the co-occurrence of the acoustic and lexical domain are relevant not only to gain a better understanding of human adaptation but may also inform the development of more naturally interacting dialogue systems.
#1312: Using Explainable AI to Identify Spanish SDOH Keywords
Ane Varela, Arantza Casillas and Maite Oronoz
Although identifying Social Determinants of Health (SDOH) is critical for health equity, datasets and tools in languages other than English remain scarce. To address this scarcity, this work presents a practical approach for the cross-lingual transfer and manual validation of SDOH factors, such as housing stability, financial stress and substance use. The experiments were carried out on a Spanish-language dataset obtained by translating the MIMIC-SBDH corpus from English, aligning the corresponding keyword annotations, and performing manual validation. Using the translated Spanish subset, we fine-tune a clinical transformer-based encoder to perform the classification task for SDOH. Then, we apply two eXplainable AI (XAI) methods, LIME and SHAP, to the classifier model and compare the generated explanations against manually validated keyword annotations. The results demonstrate that while both XAI methods capture relevant social markers, they overproduce explanations. To address this, we propose an importance weight thresholding strategy. This strategy, applied with LIME, which obtains better results overall, increases keyword overlap F1-score from 0.20 to 0.71. Furthermore, our error analysis reveals that mismatches often stem from translation errors, annotation incompleteness, and semantic overlap across classification categories. This work addresses the limited availability of resources related to SDOH in non-English settings by proposing a cross-lingual approach for transferring and evaluating SDOH information in Spanish. It also highlights the role of XAI methods assessed against a ground truth, supporting more transparent and reliable evaluation of machine learning systems.
#1337: Utterance-Level Spoken Embeddings for Dialogue State Tracking
Valentin Vielzeuf and Antoine Caubrière
Spoken Dialogue State Tracking (sDST) is typically handled with cascade systems combining ASR and text-based state tracking. An alternative is to replace transcripts with utterance-level spoken representations, which reduce input length and computational cost but may lose information needed for exact slot-value prediction. On SpokenWOZ, this approach is clearly more efficient, but it does not surpass cascade baselines under strict Joint Goal Accuracy. Fine-tuning the attention pooling layer strongly improves results, making the best utterance-level configuration competitive with a matched cascade baseline, though stronger cascade systems still perform better. Overall, the results highlight a trade-off between efficiency and value precision in utterance-level spoken DST.
#1392: UzbekSpell: An Annotated Benchmark Corpus for Error Detection and Correction in Uzbek
Botir Elov, Gulbahor Iskandarova, Xusniddin Xaitov, Dilrabo Khikmatova, Shahnoza O'dayeva, Sayyora Abdurahmanova and Aynura Akhmedova
This paper presents UzbekSpell, a controlled-access annotated benchmark corpus for detecting and correcting writing errors in Uzbek. The corpus contains 204,128 erroneous-corrected sentence pairs and 257,941 annotated error instances spanning orthographic, morphological, grammatical, punctuation, technical/coding, keyboard/OCR, and stylistic phenomena. Each instance is linked to a corrected sentence and a span-level representation of the edit. UzbekSpell was created from noisy web texts, OCR outputs, formal texts, and synthetically perturbed sentences, and annotated by five trained Uzbek linguists following shared guidelines and adjudication. To demonstrate usefulness, we benchmark a dictionary-and-edit-distance baseline against a transformer-based sequence-to-sequence correction model. On the reported test setting, the neural model substantially outperforms the baseline (P=0.82, R=0.76, F_{1}=0.79 versus P=0.60, R=0.45, F_{1}=0.51). Category-level analysis shows strongest performance on orthographic and punctuation errors, while stylistic and keyboard/OCR errors remain comparatively harder. UzbekSpell is distributed under controlled research access. To the best of our knowledge, it is the first large-scale benchmark of this breadth for Uzbek writing error detection and correction.
#1360: Your Retriever Already Knows: Predicting RAG Retrieval Sufficiency from Score Distributions
Matyáš Veselý, Michal Průšek and Jiří Franc
Standard Retrieval-Augmented Generation (RAG) pipelines often provide no reliable inference-time signal of whether retrieval succeeded; on ambiguous or out-of-scope queries, generation may then hallucinate. Motivated by a Czech nuclear-regulator deployment where data sensitivity precludes third-party LLM APIs, we compare three Query Performance Prediction (QPP) paradigms for retrieval sufficiency in RAG: score-based features, a content-based LLM judge, and a hybrid. On the public ViDoRe benchmark, GeneralQPP reaches a weighted-average AUROC of 0.856, above the Classic Full QPP baseline (0.835) and the tested local multimodal LLM judge (0.649), while remaining millisecond-level. Adding the local judge as an auxiliary feature can help in selected settings, but the gain is setting-dependent and comes at LLM latency. On the private SÚJB deployment corpus, the same ranking broadly holds; the public ViDoRe benchmark and the accompanying code release provide the reproducible basis for comparison. Our conclusions are strongest for local, privacy-constrained multimodal RAG.
