Improving speech-hearing, speech-understanding and speech-producing technologies requires supercomputers with real computing muscle – machines that can do what an ordinary computer would either do slowly or not manage at all.
"The importance of improving speech signal quality is growing," says Justina Ramonaitė, a PhD student in the Image and Signal Analysis Group at Vilnius University's Faculty of Mathematics and Informatics, Institute of Data Science and Digital Technologies (VU MIF DMSTI). "More and more technologies rely on voice, so it's important that they work properly even in noisy environments. Audio quality matters in other fields too, from communications right through to forensic expertise."
PhD student Justina Ramonaitė at the VU MIF supercomputer. Photo: Vytautas Karpauskas
Ramonaitė first used the supercomputer as a master's student, taking advantage of an opportunity VU MIF offers its students. "I was curious. One step in my research was finding the most suitable hyperparameters for the model. I decided to use the supercomputer so I could run the calculations faster."
Of everything the VU supercomputer offers researchers, the aspect most relevant to Ramonaitė is building data analysis models.
The Supercomputer Helps Models Learn Faster
Hyperparameters are settings researchers choose before training an AI model.
The model learns on its own, but it needs pre-set rules for how that learning happens: how many times it will pass through the data (researchers call each pass an "epoch"), how fast it learns, how much data it analyses at once, and how many layers deep the analysis runs.
As models grow more sophisticated, Ramonaitė says, the calculations behind them grow more complex too. Training them demands large volumes of data and many epochs. "An ordinary computer can run these calculations very slowly, or may not manage them at all for lack of resources. A supercomputer overcomes those technical barriers and speeds everything up," she says.
VU MIF's supercomputer can work with data from science, business and the public sector alike. Photo: VU MIF
Different learning settings produce different results. A high learning rate isn't necessarily better – the model may race past underlying patterns without learning them.
Too slow, and training drags on.
Too simple a system, and learning becomes inefficient, producing low-accuracy results.
Finding the best combination means testing a huge number of settings. A supercomputer can do this in parallel, trying many variants at once – so the best model turns up much faster.
Deeper Neural Networks, More Accurate Results
Ever used a rapid-translation app in a noisy train station, in a country whose language you don't speak? You've probably run into the technology's limits – the app gets lost among the overlapping sounds and conversations, and spits out a jumble of fragments.
Training speech models takes a great deal of audio – depending on the application, potentially thousands of hours – broken down into small segments across recording datasets. "With that volume of data, an ordinary computer would take weeks or months. A supercomputer can train a model in days, letting scientists and researchers test ideas and get results much faster," Ramonaitė says.
Supercomputers make it possible to train larger and more complex neural networks, potentially leading to more accurate speech recognition and synthesis models. It's not just speed that improves – systems become more accurate, sound more natural, and understand different languages better. That has real potential for assistant robots, audiobook production and translation.
Because supercomputers can train models across many languages at once, they matter for minority languages too – helping preserve linguistic diversity and widening access to technology in those languages.
PhD student Justina Ramonaitė. Photo: V. Karpauskas
"A supercomputer lets you try out more complex models than an ordinary computer could ever handle," she says.
Within the VU MIF DMSTI Image and Signal Analysis Group, Ramonaitė is currently using the supercomputer to improve speech signal quality with advanced deep learning methods, including generative adversarial networks. "If that works, we plan to explore whether feeding in information about the noise itself, present in the signal, could improve results further."
Her supervisor is Professor Dr. Gražina Korvel, who leads the development of LIEPA-3, the Large Lithuanian Language Speech Corpus.
Competing Networks
Generative adversarial networks (GANs) are an AI method for creating new, realistic-looking data – not just text, but audio and images too.
VU MIF's supercomputer. Photo: VU MIF
A GAN typically pairs two competing neural networks: a generator and a discriminator.
The generator produces fake data – an audio signal, say – while the discriminator tries to tell real data from generated data. The generator aims to fool the discriminator, the discriminator aims not to be fooled. Both improve through training – the generator's output grows more realistic, while the discriminator gets better at spotting fakes.
In audio processing, GANs generate audio signals – artificial voice recordings or music, for instance – and help augment data when real examples are scarce. They're also used to improve audio quality and reduce noise.
The Supernetwork
The high-performance computing service run by Vilnius University's Faculty of Mathematics and Informatics handles large-scale data processing. Built on open-access principles, it serves scientific research as well as business, the public sector and academic communities at other institutions.
The infrastructure forms part of the Lithuanian GRID high-performance computing network (LitGrid-HPC) and connects to the wider European EGI research computing network – a large network of scientific computers and data centres spanning Europe.
Rather than one single supercomputer, EGI links the infrastructures of many universities and institutions across different countries into a single working system. This "federation of computers" lets researchers, businesses and public bodies tap into serious computing power, analyse large datasets and collaborate internationally, even when their own institutions lack the technical capacity to do it alone.
10 August 2026
