Dutch AI Murmel Beats Major Global Models on Local Accents

A Dutch speech model is significantly better at understanding regional accents than many widely used international AI systems. In a new benchmark by Dutch company The AI Factory, the Murmel speech model made up to 30 percent fewer errors with speakers from Groningen than the best-performing open-source alternative. Limburg accents were also transcribed notably more accurately.

The results highlight a growing issue in speech tech: many AI systems handle standard Dutch reasonably well but struggle with regional pronunciation, dialect influences, and speakers outside conventional training data. According to Murmel’s developers, that creates a form of digital exclusion with real-world consequences for governments, healthcare providers, and media organizations.

Most commercial and open-source speech models are built first for English-speaking markets. Dutch is often added later with far less training data than English, making it harder for systems to recognize non-standard pronunciation patterns.

Murmel is trained specifically on thousands of hours of Dutch audio. For the benchmark, six open-source speech models were tested on nine hours of debate audio from the House of Representatives (Tweede Kamer), featuring speakers from eleven Dutch provinces.

Murmel posted the lowest error rate across every province, with the gap widening notably for regional accents.

For MPs born in Groningen, Murmel recorded a 6.4 percent error rate. The best competing model scored 9.3 percent—roughly a 30 percent improvement.

Limburg speakers were also better understood. Murmel’s error rate there was 14.6 percent, compared to 17.9 to 23.9 percent for other models. In practice, that means fewer wrong words in transcripts and far more searchable conversations.

The differences aren’t limited to regional accents. Researchers found Murmel also outperformed on MPs born outside the Netherlands.

In that group, Murmel’s error rate was 12.4 percent, while alternatives ranged from 16.1 to 20.7 percent.

According to Dr. Maarten Sukel, founder of The AI Factory and Murmel’s lead developer, inclusive AI ultimately comes down to broad intelligibility.

“Speech technology is only truly inclusive if it understands everyone—even people who aren’t perfectly clear to an AI model.”

That ties into a wider debate in AI. Many systems are trained on datasets dominated by standardized language. Regional accents, older voices, youths, and people with a migration background are often underrepresented. International studies have long shown these systems perform worse outside the norm groups they were trained on.

Murmel’s technology is already used in a range of settings: automatic transcription of meetings, city council sessions, phone calls, and healthcare conversations. Media organizations also deploy it to make radio and TV archives searchable.

According to the developers, the platform now has around 200 users, including government bodies, healthcare institutions, and media companies.

Dr. David Graus of the University of Amsterdam uses Murmel with students to transcribe city council meetings.

Graus says accurate speech recognition plays a key role in local democracy. Municipal meetings often contain information relevant to residents but, unlike national politics, receive less attention from journalists or public platforms. Automatic transcripts can make that information more accessible to citizens.

Murmel’s infrastructure is a notable differentiator. The AI Factory says all audio is processed and stored on servers in the Netherlands. Audio never has to leave the country.

That aligns with Europe’s growing focus on digital sovereignty and reducing dependence on U.S. cloud providers. European governments and companies increasingly seek AI solutions that meet GDPR requirements without relying on foreign infrastructure.

Murmel was recently listed on digitalsme.eu, a European overview of tech companies contributing to digital independence.

The model’s energy setup is also getting attention. Some GPU servers from hosting provider leaf.cloud are located in a care home in Zaandam, where their waste heat is used for hot water.

It’s a small idea, but worth noting.

Since April 2026, the infrastructure has provided roughly 297 kWh of heat, according to the company. That would displace about 34 cubic meters of natural gas and save around 250 kilograms of CO₂ compared to an average European data center.

This approach fits a broader trend: AI firms are looking for smarter ways to use the rapidly rising energy consumption of AI systems.

The rise of specialized Dutch AI models shows smaller European players are doubling down on local language models and national infrastructure. While big U.S. AI firms chase scale, Europe is building a parallel track focused on data control, privacy, and robust language support.

For languages like Dutch, that matters. Many international AI systems excel in English but lose accuracy when regional variants, accents, or local context enter the mix. In government, healthcare, and media, that can directly impact accessibility and trust.

More details on the benchmark and methodology are available on the murmel.nl website.