Two budgerigars on a branch against a dark background: one darker with dense black-and-yellow barring and a blue tail, the other a paler yellow-green

Parakeet Ultra vs Parakeet V3: How Much More Accurate Is It?

October 27, 2026(Updated October 31, 2026)
·
7 min read
·Whisper Notes Team

On September 22, 2026, Moondream released Parakeet Ultra, a further-trained version of NVIDIA's Parakeet TDT 0.6B v3. It has the same design and the same 0.6 billion parameters; only the weights changed. Moondream says it makes about 17.8% fewer mistakes across 25 languages, with the average FLEURS word error rate going from 11.62% to 9.55%. 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
English6.89%5.89%
German4.69%3.72%
Dutch7.59%6.91%
Spanish (Latin America)4.73%3.77%
French5.66%4.95%
Polish8.02%6.34%
Russian6.99%5.72%
Ukrainian6.59%5.11%
Finnish13.72%10.03%
Italian3.33%2.48%
Portuguese (Brazil)6.37%5.69%
Swedish16.18%13.12%
Czech13.70%10.85%
Lithuanian24.50%18.00%
Latvian30.25%20.70%
Greek36.10%20.96%
Estonian20.12%15.40%
Slovak11.84%8.58%
Romanian13.74%9.54%
Danish19.22%15.68%
Hungarian18.78%13.63%
Bulgarian14.05%12.28%
Croatian13.37%12.49%
Slovenian23.82%18.49%
Maltese24.46%19.19%
All 25 languages13.99%10.63%

Word error rate, the share of words transcribed wrong; lower is better. Green means Ultra is clearly better, beyond the margin of error. Croatian: no clear difference. Measured by us in Whisper Notes on an Apple M5 Mac.

Across the 25 languages, Parakeet Ultra makes about a quarter fewer mistakes than V3 (24% fewer; the word error rate drops from 13.99% to 10.63%). Where V3 got about one word in seven wrong, Ultra gets about one in nine. It was clearly better in 24 languages and worse in none.

Which languages gained the most?

Horizontal bar chart of word error rate in 25 languages for Parakeet V3 and Parakeet Ultra, sorted by the size of the drop: Greek falls from 36.10% to 20.96% and Latvian from 30.25% to 20.70% at the top, Portuguese (Brazil) from 6.37% to 5.69% at the bottom; Ultra is lower in every language, and Croatian is the only change inside its 95% interval

Word error rate by language, Parakeet V3 in gray and Parakeet Ultra in green, sorted by how far it dropped. Source: our own measurement in Whisper Notes, FLEURS test set.

The weakest languages improved the most. Latvian went 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 dropped furthest, from 36.10% to 20.96%, though it still has the highest error rate of the 25. The big languages (English, German, Spanish, French, Italian, Portuguese, Dutch and Russian) were already good and improved by a smaller 0.68 to 1.27 points, each still beyond the margin of error.

Is Parakeet Ultra slower?

Parakeet V3 Parakeet Ultra
Processing speed74.5× real time74.0× real time
Model load, median0.10 s0.10 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.

No measurable difference. Parakeet Ultra ran at 74.0× real time and Parakeet V3 at 74.5×, which is within the margin of error (speed ratio 0.99, range 0.94 to 1.06). On this Mac, an hour of audio takes about 49 seconds with either model.

That makes sense: the two models have the same design and size and differ only in their weights. Moondream describes Parakeet Ultra as built for GPUs; in Whisper Notes it runs on the Apple Neural Engine as a Core ML conversion, and it is more accurate at the same speed.

Does it hold up in background noise?

FLEURS is clean speech, so we tested noise separately. We mixed background noise from the MUSAN corpus into read sentences from LibriSpeech and from FLEURS German and Spanish, at three noise levels (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 dB4.72%3.87%−18%
LibriSpeech test-other, 5 dB5.71%4.86%−15%
LibriSpeech test-other, 0 dB9.91%8.75%−12%
FLEURS German, 10 dB9.80%8.04%−18%
FLEURS German, 5 dB13.86%10.55%−24%
FLEURS German, 0 dB24.21%17.88%−26%
FLEURS Spanish, 10 dB11.17%5.46%−51%
FLEURS Spanish, 5 dB11.90%5.77%−51%
FLEURS Spanish, 0 dB13.43%6.96%−48%
Average of the nine11.63%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.

With background noise, Parakeet Ultra made about 31% fewer mistakes than Parakeet V3 on average, and it was clearly better in all nine conditions. The noisier the German audio, the bigger the gap, and Spanish errors fell by about half at every level.

What about long recordings?

Our test used clips under 30 seconds, so it does not cover long recordings. Moondream's launch post covers long recordings with 11 full TED talks, along with English test sets, business speech and background noise:

Evaluation Parakeet V3 Parakeet Ultra
English, seven test sets6.26%5.80%
FLEURS, 25 languages11.62%9.55%
Business speech6.15%5.79%
Background noise6.72%5.82%
Long recordings, 11 TED-LIUM talks2.71%1.94%

Word error rate, lower is better; green marks the lower value. These are Moondream's own numbers: the full-precision model on Moondream's runtime and NVIDIA GPUs, with a different text normalization, so they are not comparable in absolute terms with ours above.

Details: Introducing Moondream Parakeet Redux and Parakeet Ultra

Parakeet Ultra has the lower error rate in all five; on long recordings it makes about 28% fewer errors.

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
German5.193.726.264.69
Dutch7.896.918.587.59
Slovenian20.3218.4925.5123.82
Greek34.0020.9636.4936.10
All 25 languages11.5910.6314.3013.99

Word error rate in percent. Green marks the lower value of each pair. For Parakeet Ultra every gap shown is beyond the margin of error, and 18 of the 25 languages did better with detection, none worse. We did not compute the margin for Parakeet V3 here.

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 improve the most, by 4.72 to 15.14 points.
  • English, German, Spanish, French, Italian, Portuguese and Dutch already had low error rates and improve by 0.68 to 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.
  • Chinese, Japanese and Korean are not supported by Parakeet. For those, use SenseVoice or Whisper (Whisper Large V3 Turbo on Mac, Whisper Small on iPhone).

For every model's error rate in your language, see Parakeet vs Whisper vs Qwen3-ASR vs SenseVoice.

What is Parakeet Ultra?

MakerMoondream
Base modelNVIDIA Parakeet TDT 0.6B v3
ReleasedSeptember 22, 2026
Size0.6 billion parameters, same architecture and tokenizer as the base model; about 600 MB inside the Mac app
LanguagesThe same 25 European languages as Parakeet V3
Training dataNot published
LicenseCC BY 4.0; model page: moondream/parakeet-ultra
Runs onYour device, offline, on the Apple Neural Engine, as Fluid Inference's Core ML conversion
In Whisper NotesBuilt in, no separate download, in the iPhone app and in both Mac versions (the Mac App Store version and the Mac app from our website).

We compared the converted weights with Moondream's release tensor by tensor; they match within the precision of the conversion.

How we tested it

All our FLEURS numbers were measured through Whisper Notes itself, 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 it makes about 24% fewer mistakes: the word error rate is 10.63% for Parakeet Ultra and 13.99% for Parakeet V3. Ultra was better in 24 languages and worse in none. The biggest improvements were Greek, from 36.10% to 20.96%, and Latvian, from 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, in the iPhone app and in both Mac versions. There is no separate download, and it transcribes offline on your device.

What is the difference between Parakeet Ultra and Parakeet Redux?

Both come from Moondream and start from NVIDIA's Parakeet TDT 0.6B v3. Parakeet Redux is the compressed one, for CPUs and Macs: its weights shrink from 1.2 GB to 178 MB, with the encoder's weights reduced to three possible values, and in Moondream's own tests it is a little less accurate than the original in English (6.55% against 6.26%) and clearly worse in background noise (9.04% against 6.72%). Parakeet Ultra keeps the full-precision weights and is the most accurate of the three in all five of Moondream's evaluations; it is the model Whisper Notes runs, as Fluid Inference's Core ML conversion.

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.