Fermion ResearchFermion Research

Speech recognition/facts checked 13 November 2026

Multilingual speech recognition for European languages

Accuracy across European languages varies by a factor of four for the same model. If your audio is in one language, the per-language row matters more than any average.

Phonon-2.5 downloads and benchmarks

The short answer

For the major Western European languages, open models now reach roughly 5 to 9 % word error on FLEURS, and the smaller languages run higher: Phonon-2.5 averages 11.71 % across 25 European languages, from 4.80 % in Spanish to 20.83 % in Greek. Parakeet TDT 0.6B v3 (2.5 GB) is lowest of the three models we measured in 11 languages, mostly the large ones; Phonon-2.5 (156 MB) is lowest in 6, and Parakeet Redux in the remaining 8.

Whisper covers more European languages than any of these (99 languages in total). Whisper base, which is about the size of Phonon-2.5, averages 53.71 % on the same 25 FLEURS sets.

Word error by language (FLEURS, %, lower is better)

Full FLEURS test set of each language; all three models scored with the same code and normalisation. Sizes: Phonon-2.5 156 MB, Parakeet Redux 178 MB, Parakeet TDT 0.6B v3 2,508 MB.

LanguagePhonon-2.5Parakeet ReduxParakeet TDT 0.6B v3Lowest of the three
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
Average11.7112.1813.26Phonon-2.5

Which open models cover which European languages

Language lists from each model's own card, read 2026-11-13. Coverage is not accuracy: check a per-language number before relying on a model.

ModelEuropean languagesNotes
Whisper large-v399 languages in total; the card lists themalso translates to English
Phonon-2.525: English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Russian, Ukrainian, Romanian, Czech, Slovak, Bulgarian, Croatian, Swedish, Danish, Finnish, Estonian, Latvian, Lithuanian, Hungarian, Slovenian, Maltese, Greek156 MB; transcribes only
Parakeet TDT 0.6B v3the same 25detects the language automatically, per its card
Canary 1B v225translates between English and 24 languages
ARK-ASR-3B16 of its 19, from English and German to Estonian and Lithuanianalso Chinese, Japanese, Korean
Cohere Transcribe 03-20269: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polishlanguage must be specified
Voxtral Mini 3B7 named: English, Spanish, French, Portuguese, German, Dutch, Italianalso Hindi
Granite Speech 4.1 2B5: English, French, German, Spanish, Portuguesealso Japanese

What drives the differences

  • Amount of public speech: Spanish, Italian, German and French have far more recorded speech than Maltese, Slovenian or Lithuanian, and every model in the table is more accurate on the first group.
  • Morphology: Finnish, Hungarian, Estonian and the Baltic languages build long words from many parts, so one wrong ending counts as a whole wrong word. Word error rates look worse than the transcripts read.
  • Script: Greek, Russian, Ukrainian and Bulgarian are scored in their own scripts. Greek is where the three models differ most: 20.83 % for Phonon-2.5 against 33.25 % and 36.11 % for the two Parakeet models.

Questions

Which speech recognition model is best for European languages?
It depends on the language. Among three open 25-language models measured on FLEURS with one scoring code, Parakeet TDT 0.6B v3 is lowest in 11 languages (mostly the large ones) and Phonon-2.5 in 6, with the lowest average (11.71 % against 13.26 %).
How accurate is speech recognition in German or French?
On FLEURS, Phonon-2.5 scores 6.67 % word error in German and 8.12 % in French; Parakeet TDT 0.6B v3 scores 5.96 % and 6.61 %.
Is there a small speech model for Greek, Maltese or the Baltic languages?
Phonon-2.5, 156 MB, covers Greek, Maltese, Estonian, Latvian and Lithuanian. On FLEURS it scores 20.83 % in Greek, 17.91 % in Maltese and 14.92 % in Latvian.
Does one model handle several European languages in the same file?
Phonon-2.5 and Parakeet TDT 0.6B v3 each cover all 25 languages in one model. Accuracy on files that switch language mid-sentence is not covered by FLEURS, so test on your own recordings.
Can these models translate between European languages?
Phonon-2.5 and Parakeet transcribe only. Canary 1B v2 translates between English and 24 languages, and Whisper translates into English, according to their cards.
In how many languages is Phonon-2.5 ahead of Parakeet TDT 0.6B v3?
14 of 25 on FLEURS, mostly the smaller languages. It is ahead of Parakeet Redux, a model of similar size, in 14.

Related

Sources and method

  • FLEURS numbers: Fermion Research measurement on the full FLEURS test set of each language, one scoring code and normalisation for every model (Phonon-2.5 release post, 2026-10-22).
  • Whisper base on the same 25 FLEURS sets: Phonon-2.5 release post, 2026-10-22.
  • Language lists: Hugging Face model cards of openai/whisper-large-v3, nvidia/parakeet-tdt-0.6b-v3, nvidia/canary-1b-v2, AutoArk-AI/ARK-ASR-3B, CohereLabs/cohere-transcribe-03-2026, mistralai/Voxtral-Mini-3B-2507, ibm-granite/granite-speech-4.1-2b; read 2026-11-13.

Canonical: https://www.fermionresearch.com/speech/multilingual-speech-recognition-european-languages/