Murmel on BNR: Dutch speech model outperforms big tech

Dutch business radio BNR featured Murmel in its Tech & Innovatie programme. A five-minute conversation covered why a Dutch-specific speech model outperforms the models from large international tech companies, and how Murmel is built and hosted.

Listen to the broadcast on BNR → (Dutch)

Why a Dutch model?

Models from OpenAI, Anthropic, and other large players are built to handle dozens of languages at once — English, Chinese, French, Dutch, and everything else. That breadth comes at the cost of depth: accents, dialects, and domain jargon are recognised less reliably. Murmel is trained specifically on Dutch speech, including regional accents (Haagse h, zachte g), fast speakers, and people speaking in domain-specific jargon.

Data and infrastructure

The model is trained on large volumes of Dutch audio. At some point the dataset grew large enough that additional data-centre capacity became necessary. Rather than a hyperscaler on another continent, Murmel uses the basement of a care home in Zaandam. The waste heat from the GPUs is used to warm water for residents' showers.

Technological independence

In the same week, MEPs in Strasbourg called for stronger European AI capabilities. Murmel was recently listed in the Tech Sovereignty Catalogue — a European register of digital solutions contributing to technological independence. For sectors processing sensitive audio (healthcare, government, media), this is not a detail: it means audio does not leave the Netherlands and there is no dependency on US cloud infrastructure.

Try it yourself

30 minutes of transcription included, no credit card required — or get in touch to discuss a larger deployment.