Talk about your next project
Luel content marketing

The lesson

Publish a benchmark with its losses and method caveats beside the headline result, so the claim you do make reads as tested rather than tuned.

The move

A plain header carries tags 'SPEECH SYNTHESIS' and 'DARIJA', the title 'MOROCCAN DARIJA TTS' and a 'SEE THE MODEL' button. A one-question lede asks how a dialect-native voice compares to commercial multilingual systems, and the next line admits 'Our column is best-of-8; theirs are single draws.' The left column, 'Character error · median of three native Darija recognizers', gives the method paragraph and a bar chart (Luel 0.066, ElevenLabs v2 0.088, Azure and OpenAI 0.088–0.100). The right column, 'Where we trail', lists raw audio quality at DNSMOS 3.14 against 3.37–3.45 and one judge where ElevenLabs scores lower error, and the page closes on 'The ordering is better supported than the size of the margin.'

Why it works

  • Our readOpening on a single research question tells a technical reader exactly what the benchmark does and does not test.
  • Our readThe 'Where we trail' column sits at the same level as the main result, so a skeptical reader finds the weaknesses without hunting for them.
  • Our readNaming the best-of-8 versus single-draw difference up front lets the reader judge the margin themselves, which makes the ranking easier to trust.

What not to copy

Don't copy Luel's scores, the named competitors or the Darija test setup. Keep the question lede, the one-chart result with its method stated in plain words, and a separate column that lists where you lose.

More content marketing

Steal it for your brand.

Tell us what you want to make. We’ll talk through how to put these ideas to work for your brand.