1. The 25 languages on NVIDIA’s model card
NVIDIA released parakeet-tdt-0.6b-v3 on Hugging Face on August 14, 2025, as a multilingual extension of the English parakeet-tdt-0.6b-v2. It has 600 million parameters, outputs punctuation and capitalization, and is licensed CC BY 4.0 for commercial and non-commercial use. All 25 languages share one SentencePiece tokenizer of 8,192 tokens built from the training transcripts.
The card lists the supported languages with their codes in this order: Bulgarian (bg), Croatian (hr), Czech (cs), Danish (da), Dutch (nl), English (en), Estonian (et), Finnish (fi), French (fr), German (de), Greek (el), Hungarian (hu), Italian (it), Latvian (lv), Lithuanian (lt), Maltese (mt), Polish (pl), Portuguese (pt), Romanian (ro), Slovak (sk), Slovenian (sl), Spanish (es), Swedish (sv), Russian (ru) and Ukrainian (uk).
The training mix explains the coverage. NVIDIA combined about 10,000 hours of human-transcribed speech, including LibriSpeech, Multilingual LibriSpeech, Mozilla Common Voice and Europarl-ASR, with 660,000 hours of pseudo-labeled audio from its Granary corpus, which was assembled for 25 European languages.
2. How automatic language detection behaves
The model card says Parakeet v3 automatically detects the language of the audio and transcribes it without additional prompting. The flip side is that you cannot steer it. Asked on the model’s Hugging Face page in August 2025 whether the language can be prompted or reported, an NVIDIA researcher replied on September 10, 2025 that it cannot, and that only the Canary model (canary-1b-v2) supports that. Asked about code-switching, the same researcher said the model is not fully tuned for it.
Users describe what that means in practice. One reported getting the whole transcript in English from audio that was half English and half Russian; another, dictating in Danish, saw Swedish and English words creep in; a September 2026 request for a language hint says freely spoken German with background music is regularly transcribed as English. These are individual reports, not measurements, but they point to the same habit: the model guesses the language from the audio, and noise, accents and mixed speech can tip the guess.
Whisper works differently: it also detects the language automatically, but its model card says you can pass the language explicitly when you know it in advance. If Parakeet v3 keeps choosing the wrong language for you, a Whisper model with the language fixed is the dependable fallback.
- Dictate one language per recording.
- Use a close microphone and a quiet room, since noise can tip detection.
- Switch to Whisper with a fixed language if detection keeps failing.
3. Published word error rates by language
NVIDIA’s card reports word error rate on the FLEURS test set for all 25 languages, plus Multilingual LibriSpeech (MLS) and CoVoST for some, using greedy decoding without an external language model and with punctuation and capitalization removed before scoring. The FLEURS average is 11.97%. Grouped by that FLEURS figure, the languages fall into four bands.
The other two test sets cover fewer languages. On MLS, Spanish scores 4.39%, French 4.97%, Polish 7.28%, Portuguese 7.50%, Italian 10.08% and Dutch 12.78%. On CoVoST, Russian scores 3.00%, Spanish 3.41%, Italian 3.69%, Portuguese 3.96%, German 4.84%, Ukrainian 5.10%, French 6.05%, Dutch 6.50% and English 6.80%, while Swedish (20.16%), Estonian (22.04%), Slovenian (31.80%) and Latvian (38.36%) are much higher.
Two notes from NVIDIA matter for real users. Its training data uses European Portuguese, while most benchmarks use Brazilian Portuguese, which may explain part of the Portuguese gap. And Latvian was left out of NVIDIA’s comparison chart because two of the comparison models do not support it. A 15% WER means roughly one word in seven is wrong, so for languages above that line expect corrections in most sentences and test before relying on the model.
- Under 6%: Italian 3.00%, Spanish 3.45%, Portuguese 4.76%, English 4.85%, German 5.04%, French 5.15%, Russian 5.51%.
- 6–10%: Ukrainian 6.79%, Polish 7.31%, Dutch 7.48%, Slovak 8.82%.
- 10–15%: Czech 11.01%, Romanian 12.44%, Croatian 12.46%, Bulgarian 12.64%, Finnish 13.21%.
- Over 15%: Swedish 15.08%, Hungarian 15.72%, Estonian 17.73%, Danish 18.41%, Lithuanian 20.35%, Maltese 20.46%, Greek 20.70%, Latvian 22.84%, Slovenian 24.03%.
4. German, French, Spanish and Italian against Whisper
The Hugging Face Open ASR Leaderboard’s multilingual tab, read October 3, 2026, scores both model families on the same read-speech sets (FLEURS, CoVoST-2 and MLS) for six of Parakeet’s languages. Whisper large-v3 posts a lower average WER in German (4.0% vs 4.12% for Parakeet v3), Italian (4.45% vs 4.66%), Spanish (3.35% vs 3.71%), Portuguese (4.94% vs 6.08%) and Dutch (4.42% vs 7.53%), and Parakeet v3 is ahead in French (5.38% vs 6.24%).
Against the faster Whisper large-v3-turbo, Parakeet v3 is more accurate in German (4.12% vs 6.12%), French (5.38% vs 6.73%), Italian (4.66% vs 5.43%) and Spanish (3.71% vs 3.85%), and less accurate in Portuguese (6.08% vs 4.62%) and Dutch (7.53% vs 5.70%).
Speed is where Parakeet pulls away: its multilingual RTFx on the leaderboard is 3,372, against 309 for Whisper large-v3 and 550 for turbo. For offline dictation in German, French, Spanish or Italian on an ordinary computer, that makes Parakeet v3 the practical first choice, with large-v3 worth testing when you need the last fraction of a point and can accept a slower model. In Dutch, try both.
5. Offline dictation with Parakeet v3 in IraVoice and other apps
In IraVoice, open Models, download Parakeet V3 once while you are online, and switch to it; the default Parakeet Unified EN model is English-only. From then on dictation runs on the computer with no connection. Hold Right Command on a Mac or Right Ctrl on Windows, or press Ctrl+Space on Linux, speak in any of the 25 languages, and release to have the text typed at the cursor. The license checks in weekly and keeps working for 30 days without a connection.
When Parakeet mishears a word you use often, such as a name or a place, correct it once and IraVoice learns the fix and types it correctly from then on. On a Mac it notices corrections in most apps by reading the text box it just typed into; on any platform you can press the Fix shortcut (Ctrl+Alt+Shift+F on Windows and Linux), correct the text, and press Enter. That matters most for the languages in the higher error bands above.
Other apps ship the same model. Handy, free and MIT-licensed on macOS, Windows and Linux, offers Parakeet V3 as a CPU-optimized model with automatic language detection. MacWhisper adds Parakeet v2 and v3 on M-series Macs alongside its system-wide dictation. OpenWhispr, also MIT-licensed, runs NVIDIA Parakeet through sherpa-onnx on Mac, Windows and Linux. Developers can run the model directly with NVIDIA NeMo, or with NeMo-Speech.cpp and the 8-bit GGUF file NVIDIA publishes on the model page.
- Download the model before you go offline.
- IraVoice runs on macOS 13+ on Apple silicon, Windows 10 or 11 (64-bit), and Linux (beta).
- Test a paragraph with your own names and numbers in each language you use.
6. Languages Parakeet v3 does not cover
Several European languages are missing from the 25, including Norwegian, Icelandic, Irish, Welsh, Catalan, Basque, Galician, Turkish, Serbian, Bosnian, Macedonian, Albanian and Luxembourgish. Parakeet v3 has no way to add them. On the model page, an NVIDIA researcher answered a request for Irish in October 2025 by saying NVIDIA could not find enough Irish training data, and answered a Japanese request by pointing to a separate model, parakeet-tdt_ctc-0.6b-ja; requests for Norwegian have had no answer there.
Whisper covers most of these: OpenAI’s tokenizer lists Norwegian, Nynorsk, Icelandic, Welsh, Catalan, Basque, Galician, Turkish, Serbian, Bosnian, Macedonian, Albanian and Luxembourgish, though not Irish, and it lists Japanese, Chinese, Korean, Arabic and Hindi as well. OpenAI cautions that Whisper performs unevenly across languages, with strong results in about 10, so check the language breakdown in OpenAI’s repository before relying on it.
In IraVoice, choose a Whisper model under Models for those languages. Whisper models are heavier than Parakeet, so start with Turbo or Small depending on your computer, and keep Parakeet V3 for the languages on its list.