Traditional speech therapy for children often faces challenges such as fear, a lack of patient motivation, unattractive supportive tools, and a critical shortage of specialized technological tools for under-resourced languages like Slovak. This paper presents a novel conceptual digital speech therapy ecosystem and core technological approaches developed within the 3PoCube project. To address the engagement gap, the proposed framework leverages gamification across five interconnected pillars which are integrating interactive, game-like environments for home practice with asynchronous offline telemedicine to support the parent–child–clinician dynamic. The technological core utilizes state-of-the-art Self-Supervised Learning (SSL) and Transformer-based Automatic Speech Recognition (ASR) models fine-tuned on a newly compiled Slovak children’s speech corpus, consisting of approximately 5 h of annotated speech from 303 speakers. In our experiments, the best-performing Wav2Vec 2.0 MMS architecture achieved a Word Error Rate (WER) of 15.10%, while the adapted Whisper Turbo achieved 16.33% on spontaneous pediatric speech. Several functional prototypes implementing these models are presented, including Neslovkáčik (a telemedicine portal), Hláskové superhrdinstvo (an articulation app), and LogoHra (a highly customizable virtual board game platform). Preliminary formative feedback from parents and speech-language pathologists indicates perceived acceptability during early prototype testing and suggests the potential to support engagement with the proposed solutions. However, while this study establishes a technical and conceptual foundation for home rehabilitation, rigorous clinical validation is required before definitive therapeutic efficacy can be claimed.

This paper addresses the challenge of Automatic Speech Recognition (ASR) for Slovak children’s speech, a low-resource scenario characterized by significant linguistic variability and a scarcity of dedicated corpora. We introduce the first Slovak children’s speech dataset, comprising approximately five hours of semi-automatically annotated spontaneous speech, and use it to systematically evaluate state-of-the-art ASR architectures. Specifically, we benchmark Kaldi, Whisper, and Wav2Vec 2.0 models, enriched with data augmentation techniques. Our results demonstrate that the fine-tuned Massively Multilingual Speech (MMS) Wav2Vec 2.0 model, combined with a Slovak language model, achieves a Word Error Rate (WER) of 15.10%. This represents a relative improvement of over 67% compared to our Kaldi baseline (WER $\approx ~45.44$ %). These results are highly competitive with international benchmarks for non-English child corpora, which typically report WERs in the 13.81–30.00% range. Our findings highlight three key insights: 1) dedicated data augmentation is crucial for mitigating acoustic variability in children’s speech; 2) self-supervised multilingual pretraining followed by fine-tuning yields substantial accuracy improvements; and 3) transformer-based models significantly outperform conventional hybrid ASR systems. This work establishes a strong performance baseline for future research and demonstrates the potential of modern ASR systems to support educational, therapeutic, and clinical applications for children in low-resource languages.

This paper addresses the challenge of Automatic Speech Recognition (ASR) for Slovak children's speech, a low-resource scenario characterized by significant linguistic variability and a scarcity of dedicated corpora. We introduce the first Slovak children's speech dataset, comprising approximately five hours of semi-automatically annotated spontaneous speech, and use it to systematically evaluate state-of-the-art ASR architectures. Specifically, we benchmark Kaldi, Whisper, and Wav2Vec 2.0 models, enriched with data augmentation techniques. Our results demonstrate that the fine-tuned Massively Multilingual Speech (MMS) Wav2Vec 2.0 model, combined with a Slovak language model, achieves a Word Error Rate (WER) of 15.10%. This represents a relative improvement of over 67% compared to our Kaldi baseline (WER approximate to 45.44%). These results are highly competitive with international benchmarks for non-English child corpora, which typically report WERs in the 13.81-30.00% range. Our findings highlight three key insights: 1) dedicated data augmentation is crucial for mitigating acoustic variability in children's speech; 2) self-supervised multilingual pretraining followed by fine-tuning yields substantial accuracy improvements; and 3) transformer-based models significantly outperform conventional hybrid ASR systems. This work establishes a strong performance baseline for future research and demonstrates the potential of modern ASR systems to support educational, therapeutic, and clinical applications for children in low-resource languages.

Accurate detection of land-use changes from aerial imagery is essential for urban development, environmental monitoring, and infrastructure management. While deep learning has advanced automated change detection, existing solutions remain sensitive to seasonal variations, lighting conditions, and image heterogeneity. This study presents an exploratory workflow-level evaluation of GPT-4o for semantic aerial imagery change detection and compares its behaviour with selected traditional, GIS-assisted, GAN-based, and U-Net-based approaches within the same imagery scenario. Unlike prior work focused on semantic segmentation, we investigate whether a general-purpose multimodal LLM can detect and describe changes without pixel-level training. NDVI exhibited inconsistent class separation; GAN-based mapping achieved a Structural Similarity Index Measure (SSIM) of 0.73 but lacked class fidelity; U-Net produced high accuracy for well-represented classes but struggled to generalize. GPT-4o achieved the best event-level performance, correctly identifying 89.17% of manually annotated changes and providing contextual descriptions and approximate spatial localization. Although promising, LLM performance depends on prompt specification and non-deterministic inference, raising reproducibility challenges. We address these by releasing prompt templates, raw outputs, and controlled inference settings. The results should therefore be interpreted as a case study of one proprietary multimodal model rather than as a comprehensive benchmark of all vision-language models. This exploratory study highlights the emerging potential of multimodal LLMs for interpretable and flexible geospatial analysis while outlining current limitations and future research directions.

Identifying a person's age and gender from speech signal characteristics poses a significant challenge in personal identity recognition systems, particularly when security considerations are involved. In signal processing applications such as speaker recognition, biometric identification, human-machine interface (HMI), and telecommunication, the estimation of age and gender from voice is a crucial and demanding problem. In several signal processing domains, deep learning models have demonstrated remarkable effectiveness. In this paper, we propose a modified convolutional neural network to identify the age and gender of the speaker using the characteristics of the MFCC speech. We also included techniques to reduce the dimensionality of the speech feature set. We tested modified one-dimensional convolutional neural networks (1D-CNN) and machine learning models such as support vector classification (SVC), decision trees (DT), and random forests (RF). The modified 1D-CNN based on deep learning, along with dimensionality reduction, random seeding, and cross-validation, is proposed for the recognition of age and gender in speech. We applied different dimensionality reduction techniques, such as principal component analysis (PCA) and independent component analysis (ICA), along with random seeding and various sets of cross-validation. In this study, we used the Children Speech Recorning Dataset, Biometric Visions and Computing (BVC), and the Mozilla Common Voice speech datasets for estimating age and gender from speech. The proposed 1D-CNN model exhibits a promising performance compared to the state-of-the-art (SOTA) approaches. The models were evaluated and compared with evaluation metrics, such as accuracy. The dimensionality reduction techniques, selection of speech features, and seeding show a significant impact on the performance of the suggested model.

This study presents a comprehensive evaluation and comparison of selected conventional and advanced approaches for detecting and classifying emotional stress levels from speech. Three representative and publicly available stress datasets-CRISIS, StressDat, and SUSAS-were used. A detailed comparative analysis of their speech characteristics, including fundamental frequency ( $F_{0}$ ), formant frequencies ( $F_{1}$ , $F_{2}$ , $F_{3}$ ), voice intensity, harmonic-to-noise ratio, jitter, and shimmer, was conducted to investigate how these features change with increasing stress and contribute to its detection. Both conventional feature-based methods and deep learning techniques, including transfer and self-supervised learning, are explored in the experimental part of this research. The best classification results among the conventional methods were achieved using a combination of features extracted via the Discrete Wavelet Transform and GammaTone Cepstral Coefficients (DWT+GTCC), paired with a Subspace k-Nearest Neighbors classifier, which yielded F1-scores of up to 91.1% on CRISIS and 82.4% on StressDat. Among the deep learning approaches, we fine-tuned self-supervised models such as Wav2Vec 2.0 and BYOL-S (Bootstrap Your Own Latent for Speech), as well as transfer learning models including VGGish and YAMNet. Wav2Vec 2.0 consistently outperformed the others, achieving an F1-score of 94% on CRISIS and over 87% on StressDat. On the SUSAS dataset, which contains speech under simulated and real stress conditions, an F1-score of 77% was achieved using both the fine-tuned Wav2Vec 2.0 model as well as a pipeline combining BYOL-S feature extraction with a Support Vector Machine (SVM) classifier. After merging all stress-related datasets into a single balanced dataset, Wav2Vec 2.0 achieved the highest F1-score of 85%. Incorporating additional emotional speech data improved robustness and led to a slight increase in the accuracy of classifying stress into three distinct levels. These results highlight the importance of robust feature selection, appropriate classifier design, and the integration of emotional speech data to enhance generalization performance in real-world applications.

The paper deals with the development and evaluation of short educational videos for the Object-Oriented Programming (OOP) subject to prove the role and impact of such a form of learning. A series of 17 short videos was prepared using AI-driven tools following Mayer’s principles of multimedia learning. An evaluation involving 42 OOP students through two questionnaires revealed strong positive perceptions: students found the videos useful, would use them before exams, and considered them the fastest way to acquire basic OOP knowledge. They rated short videos as the least tiring learning format compared to lectures, ChatGPT, or YouTube tutorials. The obtained findings, aligned with prior studies, demonstrate that short videos serve as an effective, low-fatigue supplement to traditional teaching, boosting motivation, accessibility, and efficiency without replacing core methods.

Children’s speech differs significantly from adult speech due to physiological and cognitive developmental factors. Key differences include higher pitch, a shorter vocal tract, greater formant frequencies, slower speaking rates, and greater variability in pronunciation and articulation. These differences result in acoustic mismatches between children’s and adult speech, making traditional automatic speech recognition models trained on adult speech less effective for children. Additionally, linguistic differences, such as limited vocabulary and evolving grammar, further contribute to this challenge. This paper focuses on the creation of a children’s speech database for the low-resource Slovak language. This database has been used to train acoustic models for the automatic recognition of spontaneous children’s speech in Slovak. In this research, we compared three different approaches to speech recognition, with self-supervised learning achieving results comparable to similar studies in this area, despite using relatively small amounts of training data.

This study addresses the issue of speech data design for the development of Romani language tests to assess the hearing abilities of preschool and primary school children in eastern Slovakia. By mapping the vocabulary of Romani words used in specific regions (Košice, Jarovnice, Bardejov), we aimed to identify the most suitable speech stimuli for audiometry. Our goal was to ensure accurate hearing evaluations by selecting words familiar to children. The work brings information directly from the course of testing and interaction with children without hearing impairment in normal conditions of educational and leisure institutions in Slovakia. We considered the linguistic differences between Romani and the majority population and emphasized the importance of native language testing for accurate hearing assessment. Based on the experience we have gained, this study contributes to the understanding of speech perception in Romani and provides practical recommendations for improving the assessment and intervention for children with hearing impairments not only in Romani-speaking communities.

Communication is crucial for humans and significantly affects their quality of life. Speech and hearing disorders result in communication and social limitations. Therefore, early diagnosis and treatment are essential. In addition to traditional methods, new supportive tools using speech analysis technologies are also helpful. This work focuses on the role of speech recognition technology in speech therapy for children with pronunciation disorders. A simple web game was designed to improve pronunciation of phonemes through children's rhymes. The game takes place on a game board where the player moves their token after correctly reciting a rhyme. The pronunciation is recorded and recognized by the system, which determines the level of agreement with the original. This score is then converted into the number of spaces the player can move. To improve the recognition of children's speech, a model was trained on a children's speech corpus, as conventional systems are optimized for adults. The newly trained model significantly improved the accuracy of speech recognition. The game increases children's motivation to practice pronunciation at home thanks to its playful elements and user-friendly interface.

The paper explores the gamification of therapeutic tools for speech and hearing disorders. The goal was to develop a web application in the form of a game to test communication development of children by the TEKOS II. The web application, game story and principles, speech and picture stimuli were designed and developed. The designed concept brings a new view of how the TEKOS test can be realized, where children are directly involved through a game-like application, which can provide additional information and which enhances the diagnostics of comprehension problems. The designed prototype can be especially useful in cases where there is no person who is able to answer questions about which words a child can understand and speak. The prototype was evaluated in short tests under laboratory conditions. The overall impression of the designed prototype is positive, where the story, speech and visual stimuli are positively assessed.

Digitization and deployment of information and communication technologies is a global trend in all areas of human activities, including the area of capturing the surrounding landscape in the form of a digital twin. These processes are also reflected in the field of spatial planning, which can effectively deploy geospatial information technologies in its processes related to its entire life cycle, from the process of gathering and evaluating suggestions, spatial planning documentations and other relevant documents to the registration and publications of the full spatial planning documentation. That is why Slovakia proceeded to create the new legislation related to spatial planning, which came into force those days and whose philosophy is based on the gradual full digitization of this issue. Part of this process is also the creation of an information system that will become a platform usable by all stakeholders. The aim of this contribution is to present the initial phase of digitization of the spatial planning process, which includes the creation of a uniform data structure for spatial plan processors, which will enable the creation of an obligatory graphic part of the spatial plan in a standardized, fully digital form.

We compare two distinct speech recognition approaches, namely Hidden Markov models mixed with deep neural networks and modern end-to-end neural speech recognition architectures. Our evaluation focuses on the metrics of performance in the context of low-resource data. Here, we utilize newly completed Slovak speech recognition dataset containing children's speech from Slovak public TV talk show intended for children. Our objective is to assess the feasibility of utilizing the Slovak children dataset within the framework of state-of-the-art end-to-end speech recognition, specifically by employing ESPnet2 framework. We anticipate the end-to-end model's performance to be inferior in comparison with any tailored hybrid Hidden Markov model. The results from the baseline experiments motivate us to employ other advanced techniques that we propose to further enhance quality of models moving forward.

The proposed paper focuses on the area of ICTbased supportive tools for speech therapy in children with speech and hearing disorders in Slovak language. The main idea was to design a concept of the supportive tool, where speech technologies can be used in a new, modern way. The web application was designed to help a child to train a correct pronunciation of particular sounds. To evaluate the similarity of the spoken word, Dynamic Time Warping algorithm was implemented, which measure distance between spoken word and the pattern in the database. Obtained distance help a caregivers or therapist to evaluate the need of continuation in speech therapy with a particular sound, or it is possible to move to another one. We also proved the concept of using automatic speech recognition in the speech therapy to create modern game-like speech therapy tools. The proposed paper brings the first observations from preliminary tests and discuss advantages and drawbacks of designed tools.

With the widespread adoption of smartphones and the increasing need for secure and personalized user experiences, user recognition has become a crucial aspect of mobile application development. In remote audiometry mobile applications, it is requested to have a quick user identification or authentication feature for protecting the medical private data. Fingerprint sensors, now commonly integrated into mobile devices, offer a convenient and reliable biometric authentication method. This paper explores using full-sized external fingerprint sensors for user recognition in mobile applications. We discuss the underlying technology of fingerprint sensors, their integration into mobile devices, and the various algorithms and techniques used for user recognition. We also present a case study where we develop a mobile application that utilizes fingerprint sensors for user authentication and discuss the performance and usability of the system.

A transition-relevance place (TRP) represents a place in a conversation where a change of speaker can occur. The appearance and use of these points in the dialogue ensures a correct and smooth alternation between the speakers. In the presented article, we focused on the study of prosodic speech parameters in the Slovak language, and we tried to experimentally verify the potential of these parameters to detect TRP. To study turn-taking issues in dyadic conversations, the Slovak dialogue corpus was collected and annotated. TRP places were identified by the human annotator in the manual labelling process. The data were then divided into chunks that reflect the length of the interpausal dialogue units and the prosodic features were computed. In the Matlab environment, we compared different types of classifiers based on machine learning in the role of an automatic TRP detector based on pitch and intensity parameters. The achieved results indicate that prosodic parameters can be useful in detecting TRP after splitting the dialogue into interpausal units. The designed approach can serve as a tool for automatic conversational analysis or can be used to label large databases for training predictive models, which can help machines to enhance human-machine spoken dialogue applications.

The paper aims to study speaker and listener behavior in dyadic speech communication. A multimodal (speech and video) corpus of dyadic face-to-face conversations on various topics was created. The corpus was manually labeled on several layers (text transcription, backchannel modality and function, POS tags, prosody, and gaze). The statistical analysis was done on the proposed corpus. We focused on backchannel inviting cues on the speaker side and backchannels on the listener side and their patterns. We aimed to study interlocutor backchannel behavior and backchannel-related signals. The results of the analysis show similar patterns in the case of backchannel inviting cues between Slovak and English data and highlight the importance of gaze direction in a face-to-face speech communication scenario. The described corpus and results of the analysis are one of the first steps leading towards natural artificial intelligence-driven human–computer speech conversation.

The proposed paper brings experiences and analyzes observations collected during children-robot spoken interaction recording during Children's University 2022 at the Technical University of Košice. We conducted the experiment in Wizard-of-Oz scenario, which was focused on backchanneling in dialogue and on some interaction scenarios, which are similar to interaction scenarios between a teacher and children during education.

The proposed paper introduces a modified Ling six sound audiometry test for the Slovak language and developed a web application, which enables checking the correct setting of the cochlear implant/hearing aid of pediatric patients. It can be used also to test hearing in the case of very young children before they are able to talk (one and half years old and older). Ling six sound test provides a set of six picture cards and related sounds, which cover the whole frequency spectrum of speech. The sounds are played sequentially, and the child selects the card that belongs to that sound. To localize the Ling test for Slovak patients the set of sounds and pictures were modified and tested. Several sets were evaluated and the pilot application was designed to be provided for further testing.

The proposed paper introduces the multimodal corpus of Slovak dyadic conversations, where backchannels from listener to speaker are labeled. The corpus consists of a series of podcasts, where two participants discuss various topics. In our research, we have focused on providing backchannels from a listener to the speaker during his/her turns. Backchannel signals play an important role in human-human spoken communication as well as in human-machine spoken dialog. They convey information about listener attention, acknowledgment, acceptance etc. The motivation for preparing the backchannel corpus in Slovak is to bring data, which can be used to train models, which will enable dialogue systems to produce backchannel signals to their human interlocutors.