Technologies
Our own AI technologies for Hungarian speech recognition, knowledge retrieval and room-acoustic design.
We develop and train the technologies and neural models below ourselves, some of them together with university researchers. They run in our own products and research projects, and kodAI and ARTIS are also available to other organisations through the Verisa platform.
kodAI
kodAI is our in-house speech recognition model, trained specifically for the Hungarian language. It is built for the way people actually speak: natural, spontaneous speech and audio carried over telephone lines. It delivers frontier-level quality in Hungarian speech recognition.
Through Verisa, our data-sovereign AI agent platform, the model can be installed on the client's own infrastructure, so audio recordings are processed within the organisation. It can be integrated into telecom operators' customer-service software and into public-administration and government systems.
- For telecom customer service: transcription of telephone calls in Hungarian
- For public administration and government bodies: on-premises speech recognition
- Installable on your own infrastructure through the Verisa platform
ARTIS
A language model's answer is only as good as the sources retrieved for it. ARTIS is the retrieval (RAG) technology we developed for the Verisa platform: for each question, it decides which parts of an organisation's documents are put in front of the model.
Common approaches cut documents into independent chunks of text and lose the connections between them along the way. ARTIS processes a document together with its structure, the way an expert would read it: chapter by chapter, keeping track of how the parts relate. This lets it answer questions that require bringing together several documents or several chapters.
It achieves frontier results on public international benchmarks. As part of the Verisa platform it runs on the customer's own infrastructure, so documents stay inside the organisation.
- Takes the structure of documents into account, not only individual chunks of text
- Gathers answers scattered across several sources
- Frontier results on public benchmarks
- On your own infrastructure, as part of the Verisa platform
FRAPPE
FRAPPE estimates a room's acoustic parameters directly from the design model: from the geometry and the material properties of the surfaces (absorption and scattering coefficients), it provides the reverberation time in each octave band. It needs no lengthy simulation, and there is no need to settle for the Sabine or Eyring formula, neither of which is reliable in complex spaces. The result is ready in milliseconds, even on a standard CPU, so even at an early stage of design it is immediately clear how a decision affects reverberation time.
The first generation of the network was built for rectangular rooms. The latest version processes the room's triangulated 3D surface model, so it also handles spaces with L-shaped, T-shaped and polygonal floor plans, sloped ceilings, columns and furnishings. Following the structure of architectural models, it works in two steps: first it summarises the geometry of each object (walls, panels, columns, furniture), then it evaluates them at room level together with their material properties.
We developed FRAPPE as part of the Soundy project, together with the artificial intelligence research group of the Institute of Mathematics at Eötvös Loránd University (ELTE).
- Octave-band reverberation time directly from the 3D model and materials
- Arbitrary floor plans, sloped ceilings, interior objects
- Response in milliseconds, even without a GPU
- Follows the structure of the architectural model: object by object first, then at room level
HERMES
Hierarchical Estimator for Room Modal Eigenvalue Synthesis
At low frequencies the sound of a room is governed by wave phenomena: room resonances (room modes). In this range, simulation based on ray tracing no longer gives reliable results, while finite-element (FEM) or boundary-element (BEM) calculations are so resource-intensive that they can rarely be run during design. HERMES covers this range as a complement to FRAPPE: it provides answers where ray-acoustic methods are no longer valid.
HERMES estimates the first 200 eigenfrequencies directly from the room's surface mesh, for rooms with rectangular, L-shaped, T-shaped and polygonal floor plans alike, including those divided by columns and objects. The network starts from physics: its estimate is based on Weyl's law, which the model refines with learned corrections. Its computational cost is orders of magnitude lower than that of conventional eigenvalue solvers and does not depend on the volume of the room.
Our sensitivity analysis shows that deviations in the sound field calculated from the HERMES estimate remain within a stable, predictable range, so the result is accurate enough for engineering design. We developed HERMES as part of the Soundy project, together with researchers from ELTE and the Budapest University of Technology and Economics (BME).
- The first 200 eigenfrequencies directly from the 3D surface model
- Physics-informed estimation built on Weyl's law, with learned corrections
- Orders of magnitude lower computational cost than conventional eigenvalue solvers
- For rooms with rectangular, L-shaped, T-shaped and polygonal floor plans, including interior objects
Would you like a demonstration of one of our technologies, or would you like to build kodAI or ARTIS into your own system? Get in touch with us.
Contact→