Image
In this success story, EuroCC Denmark helped the Danish speech recognition company Dictus ApS harness the LUMI supercomputer to train large-scale AI models for Scandinavian languages, a task that was impossible on their own hardware. The project demonstrates how European HPC resources empower Nordic SMEs to compete with global technology giants in language technology while ensuring data sovereignty and GDPR compliance.
Industrial organisations involved:
Dictus ApS is dedicated to delivering the best speech recognition technology for Nordic languages, serving clients across diverse sectors, including government, broadcasting, logistics, and healthcare. Their mission involves developing and refining speech recognition models, utilising advanced technologies like Wav2Vec and Whisper to meet specific needs such as language variations, domain-specific jargon, and varying acoustics.
Trusted by institutions such as the Danish Parliament, the Norwegian Storting, DR, TV2, PostNord, and medical professionals, Dictus consistently demonstrates expertise in speech-to-text solutions, whether for dictation workflows or the transcription of recorded audio. As demand for accurate, secure, and language-specific speech recognition continues to grow, Dictus is committed to developing solutions tailored to Scandinavian users and organisations. This includes addressing the unique linguistic characteristics of Nordic languages while ensuring high standards of quality, reliability, and data protection.
Technical/Scientific challenge:
Dictus wanted to train large end-to-end speech recognition models, such as Wav2vec and Whisper, for the Scandinavian languages. However, on their own hardware, they could only reasonably hope to train models on a thousand hours of speech. Dictus has much more data than that at the Dictus disposal - for Danish only, an order of magnitude more.
In addition, Dictus wanted to improve the robustness of the models by incorporating augmented data, speech recordings that had been synthetically manipulated to represent different acoustic conditions, speaking styles, and real-world environments. Initial experiments with data augmentation on smaller datasets had shown promising results, suggesting that the approach could significantly improve recognition accuracy and resilience. However, the computational requirements of generating and training on such large volumes made it impractical to scale using the company’s existing infrastructure.
Solution:
Training on a cluster of GPUs drastically reduced training time and made it feasible to exploit much more of the data available to Dictus, as well as the data augmentation schemes they had envisioned. Access to large-scale HPC infrastructure enabled the company to move beyond the limitations of local hardware and train significantly larger and more sophisticated speech recognition models.
Through support from EuroCC Denmark, Dictus gained guidance in accessing the LUMI supercomputer through the EuroHPC Joint Undertaking (EuroHPC JU) framework. EuroCC Denmark assisted the company throughout the application process and provided technical support during the transition of their machine learning workflows to the LUMI environment.
As part of this process, Dictus adapted its PyTorch-based training workflows from a CUDA-oriented setup to LUMI’s ROCm-based architecture, enabling efficient utilisation of the system’s GPU resources. This adaptation allowed the company’s software and training pipelines to scale effectively on one of Europe’s most powerful supercomputers while maintaining the flexibility needed for continued model development and experimentation.
Business impact:
Dictus identified a major business opportunity in offering Scandinavian speech recognition at a quality level comparable to that available for much larger languages worldwide. The company already had advanced ASR solutions in daily operation for several major customers across Scandinavia but faced strong competition from global technology companies with access to vastly greater computational resources for model training and development.
With access to the LUMI supercomputer, Dictus trained and fine-tuned ASR models on substantially larger Scandinavian datasets than previously possible, while also benefiting from advanced data augmentation techniques. Faster training cycles allowed the company to test new ideas more rapidly, improve model quality, and accelerate innovation.
The project strengthened technological capabilities in language technology and supported the development of solutions tailored to local languages and requirements.
Benefits:
• Reduced error rate in recognised text
• Fewer resources needed for post-processing and finalising text results
• Strengthened technological independence and data sovereignty
• GDPR-compliant solutions for public and private customers
• Training of larger and more robust Scandinavian language models