How machines learn human language

Allinga harnesses advanced speech recognition and synthesis to enable natural communication between humans and machines, adapted to the working environment and sector. This allows customized language solutions to be used in industry while conforming to data protection regulations.

The technology behind Allinga aims to enable machines to understand and use human language. This includes comprehending the acoustics and content of spoken commands and delivering information in a human-sounding voice. On this basis, people can work with technology in a way that feels more natural and intuitive than the current mouse, keyboard, and monitor setup. It even opens up entirely new fields of application.

Developments of this kind are already making their way into many people’s lives, for example when they control their home television or car navigation system using voice commands. Allinga is taking up this trend and laying the foundations for this technology to be used in professional settings, which have a particular need for quality, robustness, and data sovereignty. For example, if future healthcare is going to involve accessing patient data hands-free during diagnosis or operating medical devices without physical contact, physicians need to be able to trust that the technology is completely reliable and that this highly sensitive information won’t fall into the wrong hands.
 

“Made in Germany” as a guarantee of security

Although many applications can already be controlled via voice input, we have only just scratched the surface of the technological possibilities. This is where the research conducted by the Fraunhofer IIS and IAIS institutes comes in. Made in Germany, Allinga enables companies to integrate language solutions into their processes while conforming to data protection regulations. It also guarantees full data sovereignty in line with European data protection standards. To maximize security, Allinga is not reliant on the cloud. It is flexible enough to run on a company’s servers – independent of external networks, always available, and protected against unauthorized access by third parties.

Specialized AI modules for a variety of applications

Allinga is being developed by a team of more than 60 experts. Two modules – Allinga Ears for speech recognition and Allinga Voice for speech synthesis – are successfully established on the market already. They can be combined or used independently.

Allinga Voice uses a clear, fluid voice that sounds like a real person. It can be trained on local dialects or special vocabulary, allowing the system to communicate with users in the language they use every day. One feature that assists with public safety is the generation of accessible warning messages. In the event of a disaster, such as an earthquake, Allinga Voice can generate fully automated instructions for the civilian population that can be broadcast in public spaces via loudspeaker systems and on the radio – quicker and more flexibly than would be possible with human announcers. The technology enables clear, unambiguous communication while also laying the groundwork to deliver content adapted to the location in question. The application has already demonstrated its effectiveness in a simulated practical scenario: On Germany’s 2023 Nationwide Warning Day, messages generated using Allinga Voice were successfully broadcast on digital radios as part of the Emergency Warning Functionality (EWF – another technology developed at Fraunhofer IIS).

Customized configurations – even for complex working environments

Oliver Hellmuth, head of the Semantic Audio Signal Processing department at Fraunhofer IIS, explains what makes voice synthesis so special: “Allinga Voice opens up entirely new ways of using natural language: Companies can customize their language solution and adapt it to their systems and working environment. It is possible to give voices a specific accent or to develop a particular style. The technology supports multiple languages, allowing users to easily switch between German, English, and French.”

Allinga technologies offer companies numerous benefits, enabling accessible communication and natural interaction with the technology. They also feature unique properties such as modularity, data protection, openness to other technologies, and networking and distribution via an open ecosystem.

Moving into a language-based digital future together

The Allinga Voice and Allinga Ears modules are just the beginning of language technology development at Fraunhofer IIS. Its research is currently focused on Large Language Models, or LLMs for short. These AI systems can solve complex tasks given to them by humans in natural language. In the OpenGPT-X research project, several Fraunhofer institutes are working with partners from the worlds of research and industry to develop European language models that – unlike commercial offerings from the USA and China – not only adhere to the European Union’s stringent data protection regulations but have also mastered the plethora of European languages. The first product to emerge from this initiative was Teuken-7B, a large-scale AI language model released at the end of 2024. It was trained from scratch using the 24 official EU languages and encompasses seven billion parameters. Deutsche Telekom successfully transferred this development from research to business and is the first provider to offer commercial access to Teuken-7B.