Complete AI Training
Start for Free

Your job's AI kit

Get your AI kit

Tell us who you are and what you do. We show you your kit right away and email you the link: skills, prompts, AI agents, MCP servers and courses for your job.

500+ jobs ready, and we make a kit for any other job. No payment needed to look.

Share

AI news ·

Every Accent Counts: Flemish Citizens Crowdsource Speech to Train AI

KU Leuven and Scivil launch Maarallee to collect Flemish speech and cut ASR errors from NL-trained models. Record sentences and verify transcripts at maarallee.be.

Help AI catch Flemish accents: KU Leuven and Scivil launch Maarallee

KU Leuven and the Flemish Knowledge Centre for Citizen Science (Scivil) have launched Maarallee, a citizen science platform that collects speech from across Flanders to improve automatic speech recognition (ASR). The goal is simple: reduce errors caused by models trained mostly on data from the Netherlands.

If your GPS struggles with "Turnhout" in a Kempen accent, this is the fix. Record short sentences, help validate transcriptions, and move ASR closer to real Flemish speech coverage.

Why a Flemish dataset matters

Most commercial ASR models lean on Dutch data from the Netherlands. That skews pronunciation, word usage, and prosody away from Flemish speech. The result: higher error rates, frustrating user experiences, and avoidable accessibility gaps.

Broad, diverse speech data changes that. As the project partners note, without it, AI stays useful for a few and unreliable for many.

How Maarallee works

  • Contributors read short sentences on maarallee.be.
  • The platform auto-transcribes each clip; volunteers correct the text.
  • KU Leuven uses the cleaned data to build and evaluate a Flemish-focused ASR model.

Every accent, region, and age group adds signal. Even small batches of clips help close dialect gaps that standard datasets miss.

Where this matters now

  • Education and media: automatic subtitling that reflects Flemish speech.
  • Healthcare: hands-free note-taking and dictation that hear patients and clinicians correctly.
  • Industry and services: voice interfaces that work on the shop floor and in vehicles.
  • Accessibility: voice-controlled tools that respond reliably for people with disabilities.

For researchers and R&D teams

  • Expect gains in word error rate on Flemish benchmarks by expanding phonetic and dialect coverage.
  • Use the data to stress-test ASR across regions (e.g., Kempen, West-Flemish, Antwerp) and age cohorts.
  • Pair contributions with clear evaluation sets to quantify accent-specific improvements.

Participation is open to everyone. Add your voice, review a few transcripts, and help build ASR that works for all Flemish speakers.

Get involved: Record at maarallee.be. Learn more about the academic partner at KU Leuven.

Share