# No engine tops KolWrite's Oracle Max in five languages on a public test

**From the publisher · ITANU**

On the FLEURS test, no engine on the Open ASR Leaderboard has a lower word error rate than Oracle Max in five languages. In Italian it got 0.72% of the words wrong.

Software · By ITANU Magazine newsroom · Published 2026-11-22T20:01:43.000Z · 3 min read

## Key points

- We ran Oracle Max on all 4,230 clips of the FLEURS test in Italian, German, Spanish, Portuguese and French on 21 and 22 November 2026.
- Its word error rates were 0.72% in Italian, 1.58% in German, 1.76% in Spanish, 2.35% in Portuguese and 2.54% in French.
- The Open ASR Leaderboard lists 46 to 48 other engines on the same clips, and none has a lower figure in any of the five languages.

On 21 and 22 November 2026 we ran Oracle Max, the top cloud tier of KolWrite, our transcription app, on the FLEURS test in five languages: Italian, German, Spanish, Portuguese as spoken in Brazil, and French [1]. That is 4,230 recorded sentences, about 15 hours of speech [1]. The Open ASR Leaderboard publishes results for 46 to 48 other engines on the same clips [3]. In four of the languages Oracle Max has the lowest word error rate of them all, and in Spanish it shares the lowest with one other engine [1][3].

## The five results

A word error rate counts the words an engine got wrong, left out or added, as a share of the words that were spoken [1]. These are ours, each with the lowest other figure on the leaderboard for that language [1][3].

- Italian: Oracle Max 0.72%, ElevenLabs Scribe v2 0.90% [1][3]
- German: Oracle Max 1.58%, Microsoft Azure Speech 1.62% [1][3]
- Spanish: Oracle Max 1.76%, Microsoft Azure Speech 1.76% [1][3]
- Portuguese: Oracle Max 2.35%, Microsoft Azure Speech 2.45% [1][3]
- French: Oracle Max 2.54%, Microsoft Azure Speech 2.64% [1][3]

In Italian, Oracle Max wrote 753 of the 865 sentences with no error at all, and made 152 word errors in 21,223 words [1]. The leaderboard lists OpenAI's gpt-transcribe at 1.35% in Italian and 3.45% in French, and AssemblyAI's Universal-3.5 Pro at 1.03% and 2.82% [3].

## Same clips, same scoring

FLEURS is a set of sentences read aloud and recorded in 102 languages [2]. We used the test clips and the reference texts that the leaderboard uses, and scored Oracle Max with the leaderboard's own published code [1]. Before comparing anything, we rescored the leaderboard's own files with our copy of that code, and all 169 figures came out within 0.13 percentage points of its table [1].

We sent the clips through the public KolWrite API, the same service a customer uses, with the language named [1]. Every clip is counted, and none came back empty [1].

## Clip by clip

For 33 or 34 engines in each language the leaderboard publishes the transcript of every clip, so two engines can be compared on exactly the same sentences [1]. We drew 10,000 random resamples of the clips and scored both sides on each one [1]. In Italian, German and French, Oracle Max was lower than every one of those engines, with a 95% interval that leaves out zero each time [1].

In Spanish and Portuguese the same holds against 33 of the 34 [1]. Against the remaining one, Mistral's Voxtral Small, the gap is about a tenth of a point and too small to call [1]. No engine with a per-clip file was lower than Oracle Max in any of the five languages [1].

## Engines without transcripts

Microsoft Azure Speech, ElevenLabs, AssemblyAI and OpenAI appear on the leaderboard as one total for each language, with no per-clip file [1]. Against them we can set one figure beside another, and we cannot test the gap [1]. Azure's figure sits inside Oracle Max's 95% interval in German, Spanish, Portuguese and French, so those four are close [1]. In Italian the interval ends at 0.86%, below the figure of every other engine [1].

The leaderboard has further tests for these languages, and we ran FLEURS only [1]. FLEURS is clean read speech, and meetings and phone calls are harder for every engine [1]. These five are not every language we have measured: they are the ones in which no engine on this test has a lower figure [1].

## Check it yourself

Every transcript and its score are public [1]. The Italian file is at itanu.org/data/oracle-max-italian-fleurs-865-2026-11-22.csv, and the note beside it, oracle-max-five-languages-fleurs-2026-11-22.md, links the other four files and gives the method [1].

## Sources

1. [KolWrite Oracle Max on the FLEURS test in Italian, German, Spanish, Portuguese and French, measured 21 and 22 November 2026](https://itanu.org/magazine/standards/), ITANU Studios Ltd (Supplied)
2. [FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech](https://arxiv.org/abs/2205.12446), arXiv (Independent)
3. [Open ASR Leaderboard](https://huggingface.co/spaces/hf-audio/open_asr_leaderboard), Hugging Face (Independent)

ITANU’s newsroom system produced this article with AI and checked it against the sources listed above.

Canonical: https://itanu.org/magazine/kolwrite-s-oracle-max-has-the-lowest-error-rate-in-italian-yeshwcec/
