Research · Speech
How two speech systems handle Flemish news.
In a DeepTranslate test on 100 VRT news clips, about an hour of Flemish speech, Deepgram transcribed with a 9.1% word error rate and NeLF with 15.4%. The averages hide the real story: six clips caused 82% of NeLF's extra errors.
Published 14 October 2026 · DeepTranslate research
Which speech system transcribes Flemish news best?
9.1%
word error rate, Deepgram
15.4%
word error rate, NeLF
82%
of NeLF's extra errors came from just six clips
What it means for newsrooms
Average error rates hide failures. Most clips were fine for both systems. A few clips broke badly, and those are the ones viewers notice.
For broadcast, the question is not the average but the worst clip of the day.
Which errors matter most for broadcast?
| Type | Example | Why it matters |
|---|---|---|
| Missing passages | Example: Whole sentences dropped | Why it matters: Viewers miss content without knowing it |
| Names | Example: People and places misspelled | Why it matters: Errors in names cost credibility |
| Numbers | Example: 330 became 300 | Why it matters: A factual error on air |
Method and limits
- 100 publicly available VRT news clips, about one hour of speech.
- Reference transcripts are AI-generated drafts, not human-approved transcripts.
- One publisher only. Results may differ for other broadcasters.
- Clips belong to VRT. We link to the originals and do not republish video.
Frequently asked questions
Which speech recognition is most accurate for Flemish?
In our test on 100 VRT news clips, Deepgram had a 9.1% word error rate and NeLF 15.4%. NeLF still did better on 30 of the 100 clips, so neither system wins everywhere.
Can the transcripts be published without an editor?
No. Neither system was ready for unattended publishing in this test. Names, numbers and missing passages still need a human check.