Keyspoken / The voice models
Keyspoken's open Parakeet speech models, scored on accuracy, speed and noise from published benchmarks and timed on an iPhone: English, Multilingual and Ultra.
Parakeet TDT 0.6B v2
Built for English only, and the default. It is the one the app is tuned around.
6.05% word error rate
146× real time
6.95% word error rate
Parakeet TDT 0.6B v3
One model for 25 European languages. It gives up very little in English.
6.34% word error rate
156× real time
7.12% word error rate
Parakeet Ultra
Multilingual, trained further by moondream. The same 25 languages with fewer errors, in a larger download.
5.80% word error rate. Multilingual scored 6.26% on the same test.
127× real time. Multilingual ran at 129× in the same test.
5.82% word error rate. Multilingual scored 6.72% on the same test.
Scores are out of 5 and are worked out from published benchmarks. How the scores work.
Battlecard
All three side by side, on the things people rank a voice model by. Ultra is NVIDIA's Parakeet v3 trained further by moondream. We are testing it as a third choice.
6.05% word error rate
6.34% word error rate
5.80% word error rate. Multilingual scored 6.26% on the same test.
146× real time
156× real time
127× real time. Multilingual ran at 129× in the same test.
6.95% word error rate
7.12% word error rate
5.82% word error rate. Multilingual scored 6.72% on the same test.
Scores within 0.1 of each other share a place. Speed has no places: all three are too close to call.
Measured by us
The scores above come from other people's tests. These numbers are ours. We ran all three models inside the Keyspoken app on an iPhone from 2019 and one from 2021, and timed how long each took to turn speech into text.
A13 chip, 2019 · Processor
24× real time
13 seconds of speech, typed in 0.56 s with the English model.
A15 chip, 2021 · Neural Engine
132× real time
13 seconds of speech, typed in 0.10 s with the English model.
4.5 seconds of speech
13 seconds of speech
45 seconds of speech
4.5 seconds of speech
13 seconds of speech
45 seconds of speech
4.5 seconds of speech
13 seconds of speech
45 seconds of speech
Every bar uses the same scale, 0 to 2.1 seconds. Where two runs differed, the slower one is drawn.
| Model and clip | iPhone 11 Processor | iPhone 13 Pro Max Neural Engine | iPhone 13 Pro Max Processor, for comparison |
|---|---|---|---|
| English, 4.5 seconds of speech | 0.38 s | 0.08 s | 0.19 s |
| English, 13 seconds of speech | 0.56 s | 0.10 s | 0.23 s |
| English, 45 seconds of speech | 1.65 s | 0.29 s | 0.79 s |
| Multilingual, 4.5 seconds of speech | 0.41 s | 0.09 to 0.12 s | 0.20 s |
| Multilingual, 13 seconds of speech | 0.63 s | 0.18 to 0.20 s | 0.24 s |
| Multilingual, 45 seconds of speech | 1.95 s | 0.87 to 0.92 s | 0.84 s |
| Ultra, 4.5 seconds of speech | 0.41 s | 0.08 s | 0.19 s |
| Ultra, 13 seconds of speech | 0.63 s | 0.13 s | 0.25 s |
| Ultra, 45 seconds of speech | 2.01 s | 0.36 to 0.39 s | 0.87 s |
| Peak app memory | 102 to 108 MB | 69 to 84 MB | Not recorded |
| Size on the iPhone: English, Multilingual, Ultra | 443 MB, 469 MB, 612 MB | 443 MB, 470 MB, 613 MB | Same files |
How we measured. Release build. Speech clips were made with the macOS say command and bundled in the app, so the microphone was not involved. Each clip ran three times; the times shown are from the second and third runs. Measured on 12 November 2026. One phone of each kind, so treat these as what we saw, not as a promise for every iPhone.
Language by language
Word error rate on the FLEURS test set, as published by moondream for both models on the same test. Shorter bars are better. Both are strongest in Italian, Spanish, English, German, Portuguese and French.
Word error rate, FLEURS test set, as published by moondream. Sorted by Ultra, best first. Scale: 0 to 36%.
Choosing
If you only dictate in English. It is the default, it is built for English alone, and filler removal and spoken commands such as "new line" are built around it.
If you write in more than one language. It covers 25 European languages in one model, English included, and you can switch back at any time.
If you want the fewest errors across those 25 languages and have room for a larger download. It is in testing and is not in the app yet.
Method and sources
NVIDIA's published average word error rate across the eight English test sets of the Open ASR Leaderboard. A score of 5.0 means 5% or lower. Each extra percentage point costs half a point.
FluidAudio's published throughput on an Apple silicon Mac, on 5.4 hours of read English speech, in times faster than real time. A score of 5.0 means 200 times or more. Each halving costs one point.
The same word error rate, measured by NVIDIA with background noise added at a 10 dB signal-to-noise ratio, and scored the same way as accuracy.
NVIDIA did not make Ultra, so it is not in NVIDIA's tables. moondream and FluidAudio each tested it next to Parakeet v3. Ultra's score is the Multilingual score moved by the gain measured on that shared test. The figure under each bar is the published one.
What these numbers are not. The accuracy and noise figures are published results from NVIDIA and moondream, measured on data-center GPUs. The Mac speed figures are FluidAudio's. The iPhone timings are ours, measured in the Keyspoken app on an iPhone 11 and an iPhone 13 Pro Max. Keyspoken has not measured accuracy itself. The scores are ours. The figures under them are not. Tests run by different people are not directly comparable, so a gap of 0.1 between two scores is treated as a tie. An iPhone is slower than a Mac.
English uses Parakeet TDT 0.6B v2 and is the default. Multilingual uses Parakeet TDT 0.6B v3 and covers 25 European languages. Ultra is Parakeet v3 trained further by moondream for fewer errors in those same 25 languages. It is in testing and is not in the app yet. All three are open models that run on your iPhone.
If you only dictate in English, keep English. If you switch between languages, download Multilingual. You can change at any time in the app, and one model is active at a time.
The English model is about 440 MB. The app tells you the size and asks before it downloads, and you can limit it to Wi-Fi.
For English only, the English model: NVIDIA's published average word error rate across eight English test sets is 6.05% for Parakeet TDT 0.6B v2 and 6.34% for v3, the Multilingual model. Ultra, which is in testing, scored 5.80% on moondream's test where Multilingual scored 6.26%. Nobody has published English and Ultra on the same test, so those two are too close to call. None of these were measured inside Keyspoken on an iPhone.
English, on both phones we timed. With the English model, 13 seconds of speech was transcribed in 0.56 seconds on an iPhone 11 and 0.10 seconds on an iPhone 13 Pro Max. Multilingual and Ultra took 0.63 seconds on the iPhone 11, and 0.13 to 0.20 seconds on the iPhone 13 Pro Max. These are our own timings in the Keyspoken app, with speech clips rather than a live microphone.
In our own tests in the Keyspoken app, an iPhone 13 Pro Max transcribed 13 seconds of speech in 0.10 to 0.20 seconds, depending on the model. An iPhone 11, a phone from 2019, took about 0.6 seconds. On the iPhone 11 the first dictation after a cold start is slower, from 2 to 15 seconds. We timed speech clips, not a live microphone, and we have not measured accuracy.
Parakeet Ultra is NVIDIA's Parakeet v3, the Multilingual model, trained further by moondream and released under the same CC BY 4.0 licence. It covers the same 25 languages. On moondream's published tests it makes fewer errors than v3 in every one of them (9.55% average word error rate against 11.62%). It is a larger download, about 610 MB. It is in testing for Keyspoken and is not in the app yet.
The share of words a speech model gets wrong on a test recording: words it misses, adds or swaps. A 6% word error rate means about six words in a hundred differ from the reference text. Lower is better. Real results depend on your voice, your microphone and the room.
Parakeet TDT is an open speech model from NVIDIA that is small enough to run on a phone and built for transcription. Because it is open, anyone can inspect what it is. Because it runs on your iPhone, nothing you say has to travel.
It is one model trained on 25 languages. We are confirming on real devices how it handles switching languages mid-sentence before we make a promise about that.
Yes. You choose which one is active in settings. Only the active one is used for a dictation.
Not yet. Removing words like um, and commands such as new line, work in English only. With the multilingual model, your words are still transcribed and inserted as spoken.
No. Keyspoken has no way to send audio anywhere, and the models on your phone do not change as you use them.
