How can we ensure that speech technology can accurately understand the Groningen dialect, slang or a Rotterdam accent, while ensuring that the Netherlands does not become completely dependent on foreign tech companies? On 25 June, Dr Henk van den Heuvel (CLS) signed the letter of intent on behalf of Radboud University for HOSAN (High-Quality and Sovereign Speech Recognition for All Varieties of Dutch), a partnership between government, research institutions and public sector partners working towards high-quality and sovereign Dutch speech recognition.
Radar, benchmark and compass
It is precisely this dependence that matters; not the building of a speech model in itself, but the question of who will eventually do so, and with what regard for the richness of the Dutch language. "If we leave that to the market, we run the risk of our language and culture not being relevant in the AI era," says Henk.
HOSAN goes beyond building a speech model; it touches on the question of whether the Netherlands will retain control over its own language and culture in the age of AI, or whether it will remain dependent on technologies developed elsewhere by big tech, which fails to adequately address the diversity of speakers.
Science fulfils five different roles within HOSAN: a radar function, keeping an eye on international developments to Ensure a model does not become outdated before it is complete; a benchmark function, acting as an independent party that tests quality using pure test sets; a compass function, providing legal and ethical expertise for responsible data use; a bridge function, via existing national and European infrastructures; and a motor function, in the training and retention of scarce talent.
We provide the radar, the benchmark and the compass for Dutch speech recognition – Henk van den Heuvel
"HOSAN is not a research project and should not become one," Henk emphasises. "It provides a working public service; science provides the radar, the benchmark and the compass." This division of roles must take shape through a knowledge and advisory function, an independently established benchmark function, and instruments such as a PhD and graduate project programmes, internships and the shared use of infrastructure.
Dialects and slang
Specifically, Radboud University contributes expertise in the field of model evaluation, building and fine-tuning speech models, data collection and organisation, and legal frameworks such as the GDPR and the AI Act.
This is directly relevant for a problem that characterises existing speech models: they often fail to recognise dialects, regional accents and slang. "This requires the correct classification and stratification of language material, and a correct, manually verified transcription, which is essential for effective training," says Henk. "This is precisely the kind of work where linguists make a real difference."
From intention to project
The signing of the letter of intent is a first step. A CLICK-NL grant is already underway for research into better evaluation metrics for speech recognition. While the commonly used Word Error Rate primarily counts how many words have been recognised correctly, there is a need for metrics that also take into account how important a word is to understanding a sentence. The university is providing advice on this.
The focus lies on the development itself. "HOSAN is primarily a development project. We want to raise funds with major funders, such as ministries, to do something for speech that is similar to what was done with GPT-NL for language models. An alternative approach is to work with interested government partners, such as the police, to explore how we can create speech models specifically suited to their field.
Partners wishing to contribute ideas on this are welcome: the signing of the letter of intent is explicitly intended as an invitation.
Researchers behind HOSAN
Within the Faculty of Arts, there are several project in which speech recognition plays a central role, focusing on the Dutch language and all its variants. For HOSAN, Martijn Bentum is working on the technical side, building models and collecting speech recordings. Henk himself is working on the infrastructure for speech data and models, including through a one-day-a-week secondment to CLARIN-ERIC, and, drawing on his experience as a data steward at the Faculty of Arts, he contributes to the ethical and legal framework of HOSAN.