On September 22, 2026, Moondream released Parakeet Ultra, a post-trained version of NVIDIA's Parakeet TDT 0.6B v3 with the same architecture and the same 0.6 billion parameters. On Moondream's model card, its average FLEURS word error rate across 25 languages falls from 11.62% to 9.55%, about 17.8% fewer errors. We wanted to know: in real use, how much more accurate is Parakeet Ultra than Parakeet V3, language by language?
We have measured Parakeet V3, Whisper, Qwen3-ASR and SenseVoice on a fixed set of FLEURS sentences before, in our Apple SpeechAnalyzer vs Whisper benchmark; our Qwen3-ASR vs Whisper benchmark used the same set for 30 languages. This time we ran Parakeet Ultra on it. The numbers are our own, measured in Whisper Notes on one Mac, and we report them as plainly as we can.
How much more accurate is Parakeet Ultra?
We ran both models on the FLEURS test set: the 25 languages Parakeet supports, 3,976 read sentences, 12.55 hours of audio. Every sentence was transcribed the way Whisper Notes actually runs the model.
| Language | Parakeet V3 | Parakeet Ultra |
|---|---|---|
| English | 6.89% | 5.89% |
| German | 4.69% | 3.72% |
| Dutch | 7.59% | 6.91% |
| Spanish (Latin America) | 4.73% | 3.77% |
| French | 5.66% | 4.95% |
| Polish | 8.02% | 6.34% |
| Russian | 6.99% | 5.72% |
| Ukrainian | 6.59% | 5.11% |
| Finnish | 13.72% | 10.03% |
| Italian | 3.33% | 2.48% |
| Portuguese (Brazil) | 6.37% | 5.69% |
| Swedish | 16.18% | 13.12% |
| Czech | 13.70% | 10.85% |
| Lithuanian | 24.50% | 18.00% |
| Latvian | 30.25% | 20.70% |
| Greek | 36.10% | 20.96% |
| Estonian | 20.12% | 15.40% |
| Slovak | 11.84% | 8.58% |
| Romanian | 13.74% | 9.54% |
| Danish | 19.22% | 15.68% |
| Hungarian | 18.78% | 13.63% |
| Bulgarian | 14.05% | 12.28% |
| Croatian | 13.37% | 12.49% |
| Slovenian | 23.82% | 18.49% |
| Maltese | 24.46% | 19.19% |
| All 25 languages | 13.99% | 10.63% |
Word error rate, lower is better. Green: Parakeet Ultra is more accurate beyond the 95% interval. Croatian: no clear difference. Our own measurement in Whisper Notes on an Apple M5 Mac.
Pooled across the 25 languages, the word error rate fell from 13.99% to 10.63%, about 24% fewer errors: roughly one word in seven wrong became one in nine. Parakeet Ultra was more accurate in 24 languages and less accurate in none.
Which languages gained the most?
Word error rate per language, Parakeet V3 in gray and Parakeet Ultra in green, sorted by the drop in percentage points. Source: our own measurement in Whisper Notes, FLEURS test set.
The biggest gains came where Parakeet V3 was weakest: Latvian fell from 30.25% to 20.70%, Lithuanian from 24.50% to 18.00%, Slovenian from 23.82% to 18.49% and Maltese from 24.46% to 19.19%. Greek fell the most in Whisper Notes, from 36.10% to 20.96%, and is still the highest of the 25. The widely spoken languages started lower and gained less, between 0.68 and 1.27 points for English, German, Spanish, French, Italian, Portuguese, Dutch and Russian, each still outside the error bars.
Is Parakeet Ultra slower?
| Parakeet V3 | Parakeet Ultra | |
|---|---|---|
| Processing speed | 74.5× real time | 74.0× real time |
| Model load, median | 0.10 s | 0.10 s |
200 FLEURS clips, the first eight in each language, 37.2 minutes of audio. Each clip ran on both models back to back, alternating which went first. Engine processing time only, without process start or model load; not end-to-end latency in the app. Apple M5 Mac, macOS 27.0.
We found no measurable difference. Parakeet Ultra ran at 74.0× real time and Parakeet V3 at 74.5×, a speed ratio of 0.99 with a 95% interval of 0.94 to 1.06. At that rate an hour of audio takes about 49 seconds on either model on this Mac.
The two models share the same architecture and parameter count and differ only in their weights, so equal speed is what you would expect. Parakeet Ultra is more accurate at the same speed.
Does it hold up in background noise?
FLEURS is clean speech, so we checked noise in a separate test. We mixed background noise from the MUSAN corpus into read sentences from LibriSpeech and from FLEURS German and Spanish at 10, 5 and 0 dB signal-to-noise ratio: nine conditions, 1,320 sentences, run on a Mac with the same Core ML model Whisper Notes uses.
| Condition | Parakeet V3 | Parakeet Ultra | Change |
|---|---|---|---|
| LibriSpeech test-other, 10 dB | 4.72% | 3.87% | −18% |
| LibriSpeech test-other, 5 dB | 5.71% | 4.86% | −15% |
| LibriSpeech test-other, 0 dB | 9.91% | 8.75% | −12% |
| FLEURS German, 10 dB | 9.80% | 8.04% | −18% |
| FLEURS German, 5 dB | 13.86% | 10.55% | −24% |
| FLEURS German, 0 dB | 24.21% | 17.88% | −26% |
| FLEURS Spanish, 10 dB | 11.17% | 5.46% | −51% |
| FLEURS Spanish, 5 dB | 11.90% | 5.77% | −51% |
| FLEURS Spanish, 0 dB | 13.43% | 6.96% | −48% |
| Average of the nine | 11.63% | 8.02% | −31% |
Separate noise test, not the FLEURS run above: word error rate in percent, 200 LibriSpeech or 120 FLEURS sentences per condition. Its error rates are not comparable with the FLEURS table; only the relative change carries over.
Averaged over the nine conditions, Parakeet Ultra made about 31% fewer errors than Parakeet V3, and it was more accurate in all nine by more than the error bars. The gain grows with the noise in German, and Spanish improved by about half at every level.
Specifying the language makes Parakeet worse
Many Whisper Notes users have asked why Parakeet has no language setting. Parakeet can be told which language to expect, but in Whisper Notes that made it less accurate, on both versions. We measured this by running every sentence twice, once with automatic detection and once with the language specified.
| Language | Ultra, specified | Ultra, detected | V3, specified | V3, detected |
|---|---|---|---|---|
| German | 5.19 | 3.72 | 6.26 | 4.69 |
| Dutch | 7.89 | 6.91 | 8.58 | 7.59 |
| Slovenian | 20.32 | 18.49 | 25.51 | 23.82 |
| Greek | 34.00 | 20.96 | 36.49 | 36.10 |
| All 25 languages | 11.59 | 10.63 | 14.30 | 13.99 |
Word error rate in percent. Green marks the lower value within each model's pair. For Parakeet Ultra every gap shown is outside the 95% interval, and 18 of the 25 languages were more accurate with detection, none less; for Parakeet V3 we did not compute intervals.
Our guess is that specifying the language does not help Parakeet hear better. It only removes words from the transcript that don't look like that language, and names and borrowed words are often removed by mistake. So Parakeet in Whisper Notes always detects the language itself; Whisper, SenseVoice and Qwen3-ASR let you choose it, and with SenseVoice choosing it gives a slightly lower error rate.
Which languages is Parakeet Ultra good for?
- Latvian, Lithuanian, Slovenian, Maltese, Hungarian, Estonian and Greek gain the most, between 4.72 and 15.14 points.
- English, German, Spanish, French, Italian, Portuguese and Dutch already had low error rates and gain between 0.68 and 1.01 points. Croatian showed no clear difference.
- Ukrainian, Czech, Estonian, Hungarian, Lithuanian and Slovak: Parakeet Ultra now has a lower error rate than Whisper Large V3 Turbo in our FLEURS runs.
- Languages outside the 25, such as Chinese, Japanese, Korean, Arabic and Hindi: use SenseVoice Small for Chinese, Cantonese, Japanese and Korean, and Whisper for the rest of our language list, Whisper Large V3 Turbo on Mac and Whisper Small on iPhone.
What is Parakeet Ultra?
| Maker | Moondream |
| Base model | NVIDIA Parakeet TDT 0.6B v3 |
| Released | September 22, 2026 |
| Size | 0.6 billion parameters, same architecture and tokenizer as the base model; about 600 MB inside the Mac app |
| Languages | The same 25 European languages as Parakeet V3 |
| Training data | Not published |
| License | CC BY 4.0; model page: moondream/parakeet-ultra |
| Runs on | Your device, offline, on the Apple Neural Engine, as Fluid Inference's Core ML conversion |
| In Whisper Notes | Built in, no separate download: the Mac app from our website from version 1.6.4, the iPhone app from version 1.8.9. Settings lists it as “Parakeet” on Mac and “Parakeet V3” on iPhone. |
We compared the converted weights with Moondream's release tensor by tensor; they match within the precision of the conversion. Moondream, NVIDIA and Fluid Inference do not endorse Whisper Notes.
How we tested it
All numbers were measured end to end in Whisper Notes, so running the raw model elsewhere may give different results. We used the Google FLEURS test set: 25 languages, 3,976 sentences, with the same sentences and scoring as our earlier tests. Word error rate is pooled over every word in a language, 95% intervals come from resampling the sentences, and everything ran on an Apple M5 Mac. With the language specified instead, the conclusion is the same (14.30% to 11.59%), and on a separate FLEURS sample we measured the same 17.8% drop that Moondream's model card reports.
FLEURS is clean read speech in clips under 30 seconds, so meetings, far-field audio, heavy accents and long recordings are not covered; the iPhone uses the same model and was not tested separately.
Frequently Asked Questions
Parakeet Ultra vs Parakeet V3: which is more accurate?
Parakeet Ultra. Across 25 FLEURS languages in Whisper Notes its word error rate was 10.63% (Parakeet Ultra) against 13.99% (Parakeet V3), about 24% fewer errors, more accurate in 24 languages and less accurate in none. The largest gains were Greek, 36.10% to 20.96%, and Latvian, 30.25% to 20.70%.
Is Parakeet Ultra slower than Parakeet V3?
No. On an Apple M5 Mac it ran at 74.0× real time against 74.5× for Parakeet V3, a speed ratio of 0.99 with a 95% interval of 0.94 to 1.06. That is about 49 seconds for an hour of audio on either model.
What languages does Parakeet Ultra support?
The same 25 European languages as Parakeet V3: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian. For Chinese, Cantonese, Japanese and Korean, Whisper Notes offers SenseVoice Small; for other languages, Whisper.
How can I use Parakeet Ultra?
It is the built-in Parakeet model in Whisper Notes: the Mac app from our website from version 1.6.4 and the iPhone app from version 1.8.9. There is no separate download, and it transcribes offline on your device.
Try it
Parakeet Ultra is now built into Whisper Notes. Use the buttons below to get the iPhone app from the App Store or the Mac app from our website. Questions? Email support@whispernotes.app.
Download for iOS