Fermion ResearchFermion Research

Speech recognition/facts checked 13 November 2026

FLEURS results in 25 languages

The full per-language table, measured on the complete FLEURS test set of each language with one scoring code for all three models.

Phonon-2.5 downloads and model cardOpen weights on Hugging Face.

The short answer

On the full FLEURS test sets of its 25 languages, Phonon-2.5 (156 MB) averages 11.71 % word error, against 12.18 % for Parakeet Redux (178 MB) and 13.26 % for NVIDIA's Parakeet TDT 0.6B v3 (2,508 MB). All three were scored with the same code on 2026-10-22.

The average hides a split. Parakeet TDT 0.6B v3 is lower than Phonon-2.5 in 11 languages, mostly the most widely spoken ones; Phonon-2.5 is lower in the other 14. Phonon-2.5 is lower than Parakeet Redux in 14 of the 25 and the lowest of the three in 6.

Per-language word error rate

Word error rate, %, lower is better, on the full FLEURS test set of each language. Same scoring code and text normalisation for all three models; diacritics kept.

FLEURS, 25 languages, measured 2026-10-22
LanguagePhonon-2.5 (156 MB)Parakeet Redux (178 MB)Parakeet TDT 0.6B v3 (2,508 MB)Lowest
English6.587.326.53Parakeet TDT 0.6B v3
Spanish4.804.994.69Parakeet TDT 0.6B v3
French8.129.036.61Parakeet TDT 0.6B v3
German6.677.575.96Parakeet TDT 0.6B v3
Italian5.215.124.75Parakeet TDT 0.6B v3
Portuguese6.536.446.27Parakeet TDT 0.6B v3
Dutch9.219.457.93Parakeet TDT 0.6B v3
Polish10.2010.198.55Parakeet TDT 0.6B v3
Russian8.679.017.21Parakeet TDT 0.6B v3
Ukrainian7.898.057.27Parakeet TDT 0.6B v3
Romanian11.8811.8913.13Phonon-2.5
Czech10.4011.4111.93Phonon-2.5
Slovak8.898.179.72Parakeet Redux
Bulgarian13.3413.0213.48Parakeet Redux
Croatian12.2411.4313.21Parakeet Redux
Swedish14.5314.7015.44Phonon-2.5
Danish18.2017.9119.05Parakeet Redux
Finnish12.0211.8413.08Parakeet Redux
Estonian11.9110.2914.34Parakeet Redux
Latvian14.9214.7623.66Parakeet Redux
Lithuanian17.9418.0521.68Phonon-2.5
Hungarian15.4515.8115.38Parakeet TDT 0.6B v3
Slovenian18.5318.8824.70Phonon-2.5
Maltese17.9116.0220.86Parakeet Redux
Greek20.8333.2536.11Phonon-2.5
Average of 2511.7112.1813.26Phonon-2.5

Where each model is ahead

  • Parakeet TDT 0.6B v3 is lower than Phonon-2.5 in: English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Russian, Ukrainian, Hungarian. If your audio is mostly in these languages and a 2.5 GB model fits your hardware, it is the more accurate choice of the three.
  • Parakeet Redux is lower than Phonon-2.5 in: Italian, Portuguese, Polish, Slovak, Bulgarian, Croatian, Danish, Finnish, Estonian, Latvian, Maltese.
  • Phonon-2.5 is the lowest of the three in: Romanian, Czech, Swedish, Lithuanian, Slovenian, Greek.
  • Greek is the largest gap: Phonon-2.5 20.83 %, Redux 33.25 %, Parakeet v3 36.11 %. Leave Greek out and Phonon-2.5 and Redux average 11.34 % and 11.31 % over the other 24, close to a tie; Parakeet v3 averages 12.31 %.

Reading the table

Error rates vary far more by language than by model. Phonon-2.5 is under 10 % in 10 languages and at 15 % or more in 6, from 4.80 % in Spanish to 20.83 % in Greek. Languages with rich word forms, such as Finnish, Hungarian or the Baltic languages, tend to score higher on a word-level metric because one wrong ending costs a whole word.

FLEURS is read speech. Spontaneous speech, phone audio and meetings are harder, and these numbers are not a promise for that audio. Because our scoring keeps diacritics, these figures are not comparable with FLEURS numbers published elsewhere with other normalisers.

For scale on the same 25 sets and the same scoring: Whisper base, about the size of Phonon-2.5, averages 53.71 %.

Questions

What is Phonon-2.5's average FLEURS word error rate?
11.71 % across its 25 languages, on the full FLEURS test set of each language, scored with one code for all models compared.
Which languages does the FLEURS table cover?
The 25 languages of Phonon-2.5: English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Russian, Ukrainian, Romanian, Czech, Slovak, Bulgarian, Croatian, Swedish, Danish, Finnish, Estonian, Latvian, Lithuanian, Hungarian, Slovenian, Maltese and Greek.
Is Phonon-2.5 more accurate than Parakeet TDT 0.6B v3?
On average over the 25 languages, yes (11.71 % against 13.26 %). Parakeet TDT 0.6B v3 is lower in 11 languages, including English, French, German and Spanish.
Which language has the highest word error for Phonon-2.5?
Greek, at 20.83 %. Greek is also where Phonon-2.5 is furthest ahead of the other two models.
Why are these numbers different from FLEURS results in papers?
Scoring choices differ. Our code keeps diacritics and is the same for every model in the table; papers that strip diacritics or use other references get different numbers on the same audio.
How big are the three models?
Phonon-2.5 is a 156 MB download, Parakeet Redux 178 MB and Parakeet TDT 0.6B v3 2,508 MB.

Related

Sources and method

  • FLEURS numbers: Fermion Research measurement, published 2026-10-22, full FLEURS test set per language, one scoring code and normalisation for all three models, diacritics kept.
  • Whisper base on the same 25 FLEURS sets: Phonon-2.5 release post, 2026-10-22.
  • FLEURS dataset: huggingface.co/datasets/google/fleurs, read 2026-11-13.

Canonical: https://www.fermionresearch.com/speech/fleurs-results-25-languages/