As the output quality of neural networks in the fields of automatic speech recognition (ASR) and text-to-speech (TTS) continues to improve, new opportunities are becoming available to train models in a weakly supervised fashion, thus minimizing the manual effort required to annotate new audio data for supervised training. While weak supervision has recently shown very promising results in the domain of ASR, speech synthesis has not yet been thoroughly investigated regarding this technique despite requiring the equivalent training dataset structure of aligned audio-transcript pairs.
In this work, we compare the performance of TTS models trained using a well-curated and manually labeled training dataset to others trained on the same audio data with text labels generated using both grapheme- and phoneme-based ASR models. Phoneme-based approaches seem especially promising, since even for wrongly predicted phonemes, the resulting word is more likely to sound similar to the originally spoken word than for grapheme-based predictions.
For evaluation and ranking, we generate synthesized audio outputs from all previously trained models using input texts sourced from a selection of speech recognition datasets covering a wide range of application domains. These synthesized outputs are subsequently fed into multiple state-of-the-art ASR models with their output text predictions being compared to the initial TTS model input texts. This comparison enables an objective assessment of the intelligibility of the audio outputs from all TTS models, by utilizing metrics like word error rate and character error rate.
Our results not only show that models trained on data generated with weak supervision achieve comparable quality to models trained on manually labeled datasets, but can outperform the latter, even for small, well-curated speech datasets. These findings suggest that the future creation of labeled datasets for supervised training of TTS models may not require any manual annotation but can be fully automated.

There are many publications talking about the biases to be found in in generative AI solutions like large language models (LLMs, e.g., Mistral) or text-to-image models (T2IMs, e.g., Stable Diffusion). However, there is merely any publication to be found that questions what kind of behavior is actually desired, not only by a couple of researchers, but by society in general. Most researchers in this area seem to think that there would be a common agreement, but political debate in other areas shows that this is seldom the case, even for a single country. Climate change, for example, is an empirically well-proven scientific fact, 197 countries (including Germany) have declared to do their best to limit global warming to a maximum of 1.5°C in the Paris Agreement, but still renowned German scientists are calling LLMs biased if they state that there is human-made climate change and humanity is doing not enough to stop it. This trend is especially visible in Western individualistic societies that favor personal well-being over common good. In this article, we are exploring different aspects of biases found in LLMs and T2IMs, highlight potential divergence in the perception of ethically desirable outputs and discuss potential solutions with their advantages and drawbacks from the perspective of society. The analysis is carried out in an interdisciplinary manner with the authors coming from as diverse backgrounds as business information systems, political sciences, and law. Our contribution brings new insights to this debate and sheds light on an important aspect of the discussion that is largely ignored up to now.

Vergnügungsparks stehen vor der Herausforderung, sich in einer zunehmend digitalisierten Welt weiterzuentwickeln, um ihre Attraktivität für ein breites Publikum zu sichern. Digitale Technologien, insbesondere Virtual Reality (VR), eröffnen hier neue Perspektiven. Sie verändern nicht nur das Besuchererlebnis, sondern haben auch einen signifikanten Einfluss auf den Geschäftserfolg.

In this paper we present a stable-diffusion based zero-shot approach to realistically transform the image of a

human body into a more fit version of that depicted person. Therefore we combine a modified stable diffusion

model with inpainting techniques and incorporated constraints. We introduce a prototype which allows users to

upload a photo and visualize a more fit version of themselves. We evaluated our approach in various experiments

and focused on the applicability and effectiveness of these techniques, with attention to gender-specific results.

This work contributes to the fields of computer vision and generative AI by demonstrating practical applications

and identifying areas for improvement in realistic body transformation visualizations. (This work is a part of the project (M4-SKI) has been supported and funded by the European Regional Development Fund (ERDF)).

Natürlich sollte man nichts verkomplizieren. Oft sind ja einfache Lösungen durchaus sinnvoll. Aber angesichts der Komplexität unserer Umwelt neigen wir offenbar dazu, Probleme auf vermeintlich eindeutige Ursachen zurückzuführen. Dieser „Simplification Bias“ bestimmt zunehmend den gesellschaftlichen Diskurs, führt aber auch in Managementfragen zu schlechten Entscheidungen.

Loyalty Marketing hat sich in den letzten Jahren stark gewandelt. Unternehmen müssen heute weit mehr tun, als nur Rabattkarten auszustellen oder Treuepunkte zu vergeben, um ihre Kunden langfristig zu binden. Künstliche Intelligenz (KI) spielt dabei zunehmend eine zentrale Rolle, insbesondere bei der Datenanalyse, Automatisierung und Personalisierung.

Background

The impact of the infecting SARS-CoV-2 variant of concern (VOC) and the vaccination status was determined on the magnitude, breadth, and durability of the neutralizing antibody (nAb) profile in a longitudinal multicentre cohort study.

Methods

173 vaccinated and 56 non-vaccinated individuals were enrolled after SARS-CoV-2 Alpha, Delta, or Omicron infection and visited four times within 6 months and nAbs were measured for D614G, Alpha, Delta, BA.1, BA.2, BA.5, BQ.1.1, XBB.1.5 and JN.1.

Findings

Magnitude-breadth-analysis showed enhanced neutralization capacity in vaccinated individuals against multiple VOCs. Longitudinal analysis revealed sustained neutralization magnitude-breadth after antigenically distant Delta or Omicron breakthrough infection (BTI), with triple-vaccinated individuals showing significantly elevated titres and improved breadth. Antigenic mapping and antibody landscaping revealed initial boosting of vaccine-induced WT-specific responses after BTI, a shift in neutralization towards infecting VOCs at peak responses and an immune imprinted bias towards dominating WT immunity in the long-term. Despite that bias, machine-learning models confirmed a sustained shift of the immune-profiles following BTI.

Interpretation

In summary, our longitudinal analysis revealed delayed and short lived nAb shifts towards the infecting VOC, but an immune imprinted bias towards long-term vaccine induced immunity after BTI.

Funding

This work was funded by the Bavarian State Ministry of Science and the Arts for the CoVaKo study and the ForCovid project. The funders had no influence on the study design, data analysis or data interpretation.

Der Rückgriff auf ChatGPT & Co. vereinfacht vieles im Alltag. Es ist unkompliziert und naheliegend, sich insbesondere Texte durch generative KI schreiben zu lassen oder auch Zusammenfassungen von komplexen und langen Artikeln damit zu erstellen, gerade wenn Zeit und Aufmerksamkeit begrenzt sind. Aber werden wir damit nicht zu bequem? Lassen wir unsere grauen Zellen damit verkümmern?Oder verkennen wir mit solchen Fragen das Potenzial der Technologie? Und welche Kompetenzen brauchen wir dann überhaupt in Zukunft noch?

This paper investigates the comparative effectiveness of model-to-model transformations generated by an LLM based upon user prompts versus those created with dedicated model transformation languages, using a standard benchmark. The emergence of Generative AI offers a novel approach, allowing developers to specify transformations in natural language rather than learning specialized languages. However, our findings suggest that, in its current state, generative AI does not yet pose a threat to dedicated model transformation languages. While AI-assisted approaches promise to provide flexibility and accessibility, dedicated model transformation languages still offer structured advantages crucial for complex transformations, especially when bidirectionality and incrementality are mandatory requirements. This research contributes to the ongoing discourse on the role of AI in software engineering, highlighting its potential and current limitations in enhancing model transformation processes.

Being able to maintain the consistency between various different, but related models is an important enabler for model-based software engineering. Research on bidirectional transformations (bx) addresses this issue and has resulted in various and diverse formal foundations, approaches, tools, and application scenarios. In order to understand and compare different bx approaches, we have developed benchmarx, a benchmarking framework specifically for bx. Up until now, however, benchmarx has been limited to one-sided model synchronisation tasks, where only one of two related models can be changed at a time.As the more general case of concurrent model synchronisation is crucial for many practical applications of bx, we propose in this paper an extension to our bx benchmarking framework to support concurrent model synchronisation tasks, where two related models can both be changed concurrently and must then be synchronised to restore consistency. To evaluate our new extensions we present an update of an existing benchmarx example, families-to-persons, to include new test cases requiring concurrent synchronisation. We discuss some of the challenges involved in defining such a benchmark including handling conflicts, defining the expected behavior of the bx tool under test, and providing bx tools with enough freedom to reject some of the changes to either model. We also present a solution to the updated families-to-persons benchmarx example implemented using BXtend as a bx tool.