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.

Dementia is a growing global health challenge, making early detection of cognitive decline critically important. Early detection of dementia and other cognitive impairment is essential for timely intervention and better care planning. However, existing datasets for training automated screening tools are limited, especially for underrepresented languages such as Czech. In this study, we present a new dataset and a novel application named the DigiDiaDem (Digital Diagnostics of Dementia) designed for automated dementia screening through multimodal cognitive assessment. The application integrates user-friendly digital cognitive tasks with machine learning algorithms to evaluate linguistic and cognitive performance in real time. Using this system, we collected and curated a comprehensive dataset of speech and cognitive data from Czech-speaking participants. The dataset comprises 371 individuals, including cognitively normal individuals and patients with mild cognitive impairment and mild dementia. It includes socio-demographic data, results of cognitive and speech tests, functional assessment questionnaires, data collected through the DigiDiaDem application, and automatic speech recognition (ASR) transcripts of spoken responses. Raw audio recordings are not included. Instead, the dataset provides manually engineered linguistic and acoustic features. We describe the data collection process and outline the cognitive tasks used to collect the dataset. Our experiments demonstrate that speech features derived from cognitively demanding tasks, such as verbal fluency and memory recall, can effectively distinguish healthy participants from those with cognitive impairment. Models trained on the dataset achieved up to 95% accuracy when combining speech features with demographic information. Preliminary experiments demonstrate the feasibility of using the collected data for dementia detection. These findings confirm that speech-based digital assessment can complement traditional clinical evaluation. The proposed dataset and application offer a substantial resource for the research community by establishing a solid baseline for machine-learning-based approaches to dementia screening from speech-based interaction.

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.

The rapid advancement of pre-trained large language models (LLMs) has enabled the creation of innovative applications, especially in natural language processing. This work employs LLMs alongside our in-house technologies to develop an intuitive database search engine that processes natural language queries. The system uses a network of AI agents, including prompted LLMs and single-purpose neural classifiers, to categorize user queries into conditions for filtering individual data sources or direct matches to database entries. Enhanced with a Retrieval-Augmented Generation (RAG) approach, the application allows users to search large databases conversationally through a voice-enabled web-based interface. Currently, in the demo stage, this project shows full pipeline functionality and has been tested with approximately 150 h of transcribed speech data. Initial findings confirm the overall concept of the application.

We investigate whether immediate repetition of nonsense words can distinguish cognitively healthy adults from those with mild cognitive impairment (MCI) or dementia. In a computer-based study, 129 Czech speakers (45–84 y) repeated six pseudowords; each session was recorded and transcribed by four state-of-the-art ASR models. Grapheme-level similarity between the transcript and the target word served as a phonological accuracy score. Using logistic regression, the best ASR variant (wav2vec-nolm, no language model) separated patients from controls with 77

Automatic prediction of mean-opinion scores (MOS) promises a faster, cheaper alternative to listening tests, yet robust generalization across speakers, languages, and domains remains a significant challenge. This article presents our system designs and experimental results from three years of participation in the VoiceMOS Challenges (2022-2024), covering MOS prediction for synthesized or voice-converted speech and singing voice, including out-of-domain and cross-language conditions. We evaluate six neural architectures - wav2vec 2.0, QuartzNet, CNN-RNN, LDNet, RawNet3, and HiFi-GAN - and their ensembles. Across all tasks, we find that 1) self-supervised acoustic encoders are the most consistently reliable foundation, 2) ensembling yields rapidly diminishing returns once complementary representations are covered, and 3) the diversity and balance of training data outweigh architectural complexity. Notably, the indiscriminate fusion strategy that performed well in 2022 degrades under the mismatched French TTS conditions of 2023, emphasizing the importance of out-of-domain validation. Further experiments show that carefully pruned ensembles can modestly outperform the best single model while remaining within real-time constraints. We conclude with several observations to guide the development of computationally efficient, domain-robust MOS prediction systems.

Background: The preservation and analysis of spoken data in oral history archives, such as Holocaust testimonies, provide a vast and complex knowledge source. These archives pose unique challenges and opportunities for computational methods, particularly in self-supervised learning and information retrieval.Objective: This study explores the application of state-of-the-art artificial intelligence (AI) models, particularly transformer-based architectures, to enhance navigation and engagement with large-scale oral history testimonies. The goal is to improve accessibility while preserving the authenticity and integrity of historical records.Methods: We developed an asking questions framework utilizing a fine-tuned T5 model to generate contextually relevant questions from interview transcripts. To ensure semantic coherence, we introduced a semantic continuity model based on a BERT-like architecture trained with contrastive loss.Results: The system successfully generated contextually relevant questions from oral history testimonies, enhancing user navigation and engagement. Filtering techniques improved question quality by retaining only semantically coherent outputs, ensuring alignment with the testimony content. The approach demonstrated effectiveness in handling spontaneous, unstructured speech, with a significant improvement in question relevance compared to models trained on structured text. Applied to real-world interview transcripts, the framework balanced enrichment of user experience with preservation of historical authenticity.Conclusion: By integrating generative AI models with robust retrieval techniques, we enhance the accessibility of oral history archives while maintaining their historical integrity. This research demonstrates how AI-driven approaches can facilitate interactive exploration of vast spoken data repositories, benefiting researchers, historians and the general public.

Timely detection of cognitive disorders is critical yet often limited by resource-heavy clinical assessments. This paper presents an automated framework for early cognitive screening based on the semantic analysis of spoken image descriptions in Czech. The system integrates automatic speech recognition, formal semantic parsing, and machine learning to evaluate deviations from an expert-defined reference description. The responses of the participants are analyzed for missing or incorrect semantic content, producing structured loss vectors used for classification. Evaluation on a clinically annotated dataset of 268 samples (split into train-test subsets) shows that semantic features outperform traditional lexical and morphological baselines, highlighting the potential of the method for scalable and interpretable cognitive assessment.

The importance of oral history archives in preserving and understanding past experiences is counterbalanced by the challenges encountered in accessing and searching through them, primarily due to their extensive size and the diverse demographics of the speakers. This paper presents an approach combining ASR technology and Transformer-based neural networks into the Asking questions framework. Its primary function is to generate questions accompanied by concise answers that relate to the topics discussed in each interview segment. Additionally, we introduce a semantic continuity model that filters the generated questions, ensuring that only the most relevant ones are retained. This enables a real-time semantic search through thousands of hours of recordings, with the crucial benefit that the speakers' original words remain unaltered and still semantically align with the query. While the method is exemplified using a specific publicly available archive, its applicability extends universally to datasets of a similar nature.

This paper describes our team's experiences in the VoiceMOS Challenge 2023 - a challenge centered around the evaluation of the quality of synthetic or noisy speech. Inspired by our success with an ensemble approach in the first VoiceMOS Challenge in 2022, we submitted an ensemble of four models this time, based on wav2vec 2.0, QuartzNet, CNN-RNN, and LDNet. This was enough to win one of the two tracks we participated in (Track 1b). However, post-challenge analysis shows that only two of the models offer a meaningful contribution in any of the VoiceMOS 2023 tracks, while the other two only degrade the ensemble's overall performance. On the other hand, post-challenge results on Track 2 (singing voice conversion data) surpassed all our expectations. In the paper, we explain how we tried to deal with the new zero-shot out-of-domain scenarios, analyze the results, and discuss the lessons learned.

Implementing AI solutions developed in controlled lab environments into practical scenarios reveals challenges such as hardware constraints or unpredictable real-world conditions. This paper introduces a cost-effective robotic entity designed exclusively for validating human-machine interaction tasks in real-life scenarios. Featuring flexible visual, auditory, and other sensory interfaces, the device serves as a platform for deploying and validating students as well as advanced projects in AI. It effectively bridges the gap between theory and practice, embodying generality and multimodality. Ongoing efforts demonstrate practicality through successfully implemented applications in the field of human-machine interaction. Copyright (c) 2024 The Authors.

The paper presents a method for spoken term detection based on the Transformer architecture. We propose the encoder-encoder architecture employing two BERT-like encoders with additional modifications, including attention masking, convolutional and upsampling layers. The encoders project a recognized hypothesis and a searched term into a shared embedding space, where the score of the putative hit is computed using the calibrated dot product. In the experiments, we used the Wav2Vec 2.0 speech recognizer. The proposed system outperformed a baseline method based on deep LSTMs on the English and Czech STD datasets based on USC Shoah Foundation Visual History Archive (MALACH).

Abstract Although Czech rule-based tools for automatic punctuation insertion rely on extensive grammar and achieve respectable precision, the pre-trained Transformers outperform rule-based systems in precision and recall (Machura et al. 2022). The Czech pre-trained RoBERTa model achieves excellent results, yet a certain level of phenomena is ignored, and the model partially makes errors. This paper aims to investigate whether it is possible to retrain the RoBERTa language model to increase the number of sentence commas the model correctly detects. We have chosen a very specific and narrow type of sentence comma, namely the sentence comma delimiting vocative phrases, which is clearly defined in the grammar and is very often omitted by writers. The chosen approaches were further tested and evaluated on different types of texts.

The paper introduces software capable of indexing and searching large archives of scanned historical documents. The system capabilities are demonstrated on the collection containing documents from the archives of the post-Soviet security services. The backend of the system was designed with a focus on flexibility (it is actually already being used for other related tasks) and scalability to larger volumes of data. The graphical user interface design has been consulted with historians interested in using the archived documents and was developed in several iterations, gradually including the changes induced both by the user’s requests and by our improving knowledge about the nature of the processed data.

This paper is a step forward in our effort to make vast oral history archives more accessible to the public and researchers by breaking down the decoding barriers between the knowledge encoded in the spoken testimonies and users who want to search for the information of their interest. We present new Transformer-based monolingual models suitable for speech recognition of oral history archives in English, German, and Czech. Our experiments show that although the all-purpose speech recognition systems have recently made tremendous progress, the transcription of oral history archives is still a challenging task for them; our tailored models significantly outper-formed larger public multilingual models and scored new state-of-the-art results on all tested datasets. Due to the 2-phase fine-tuning process, our models are robust and can be used for oral history archives of various domains. We publicly release our models within a public speech recognition service.

The paper describes our initial effort to use Transformer-based neural networks for understanding and presenting oral history archives. Such archives of interviews often contain large passages of the interviewee's speech. Our approach automatically generates relevant questions, which enrich such monotonous parts and allows the listener to better orient in the interview. The generated questions also allow for finding interesting parts of the interview without changing the original meaning of the testimony. We present our working pipeline consisting of a Wav2Vec speech recognizer, BERT-based punctuation detection, T5 asking questions model and BERT-based semantic continuity model.

Traditional offline learning approaches are reaching their limits in meeting the dynamic demands of specialized applications, such as real-time human-robot interaction. While high benchmark scores attained through offline fine-tuning large models on extensive data, offer a glimpse of their potential, the true functionality is validated when these models are deployed on target devices and utilized in real-life scenarios. This paper presents a method incorporating humans in an interactive learning loop, using their real-time feedback for online neural network retraining. By leveraging the power of transfer learning, we can proficiently adapt the model to suit the specific requirements of the target application through natural voice-based dialogue. The approach is evaluated on the image classification task utilizing a unique low-cost device and a practical example of the real-time dialogue is presented to demonstrate the functionality.

This paper discusses the analogies between the mainstream theory of human mind and the two broad paradigms that are employed when building artificial intelligence systems. Then it ponders the idea how those analogies could be utilized in building a truly explainable artificial intelligence (AI) applications. The core part is devoted to the problem of unwanted rationalization that could disguise the true reasons lying behind the decisions of the explainable AI systems.