# The voice models

> Keyspoken. Markdown copy of https://keyspoken.com/models. Publisher: Kymata Labs LLC. Last updated: 2026-11-12.

- English: Parakeet TDT 0.6B v2, by NVIDIA. Languages: English. Default, downloaded at setup. Runs on the device. Filler removal and spoken commands.
- Multilingual: Parakeet TDT 0.6B v3, by NVIDIA. Languages: 25 European languages, English included. Optional download. Runs on the device. Filler removal and spoken commands in English only.
- Ultra: Parakeet Ultra, by moondream, trained further from NVIDIA's Parakeet v3. Languages: 25 European languages, English included. In testing, not in the app yet. Runs on the device. Filler removal and spoken commands in English only.

Languages covered by the multilingual model: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish, Ukrainian.

### Published benchmarks

- Parakeet TDT 0.6B v2 (English; the default in Keyspoken). Accuracy: 6.05% word error rate. Speed on an Apple silicon Mac: 146× real time. With background noise: 6.95% word error rate. Keyspoken scores out of 5: accuracy 4.5, speed 4.5, noise 4.0. Licence: CC BY 4.0.
- Parakeet TDT 0.6B v3 (Multilingual; an optional download in Keyspoken). Accuracy: 6.34% word error rate. Speed on an Apple silicon Mac: 156× real time. With background noise: 7.12% word error rate. Keyspoken scores out of 5: accuracy 4.3, speed 4.6, noise 3.9. Licence: CC BY 4.0.
- Parakeet Ultra (Ultra; in testing for Keyspoken, not in the app yet). Accuracy: 5.80% word error rate. Multilingual scored 6.26% on the same test. Speed on an Apple silicon Mac: 127× real time. Multilingual ran at 129× in the same test. With background noise: 5.82% word error rate. Multilingual scored 6.72% on the same test. Keyspoken scores out of 5: accuracy 4.6, speed 4.6, noise 4.4. Licence: CC BY 4.0.

Word error rate by language on FLEURS, as published by moondream, Multilingual then Ultra: Italian 2.6% and 2.0%; Spanish 3.1% and 2.7%; English 4.3% and 3.5%; German 4.1% and 3.6%; Portuguese 4.7% and 4.0%; French 4.8% and 4.3%; Ukrainian 5.9% and 4.8%; Russian 5.9% and 5.2%; Dutch 6.2% and 5.5%; Polish 6.7% and 5.5%; Slovak 9.5% and 7.0%; Romanian 11.5% and 9.2%; Finnish 11.1% and 9.2%; Croatian 10.9% and 9.7%; Estonian 13.2% and 9.7%; Czech 10.8% and 10.0%; Bulgarian 11.9% and 10.1%; Hungarian 13.7% and 10.8%; Swedish 13.8% and 11.6%; Danish 16.8% and 14.3%; Maltese 19.1% and 14.9%; Latvian 21.4% and 15.9%; Lithuanian 21.1% and 16.4%; Slovenian 21.8% and 16.6%; Greek 35.7% and 32.3%.

### Timed by us on two iPhones

- iPhone 11 (A13 chip, 2019, Processor): English 4.5 s of speech in 0.38 s, 13.2 s of speech in 0.56 s, 45 s of speech in 1.65 s; Multilingual 4.5 s of speech in 0.41 s, 13.2 s of speech in 0.63 s, 45 s of speech in 1.95 s; Ultra 4.5 s of speech in 0.41 s, 13.2 s of speech in 0.63 s, 45 s of speech in 2.01 s. Peak app memory 102 to 108 MB. Cool and at rest.
- iPhone 13 Pro Max (A15 chip, 2021, Neural Engine): English 4.5 s of speech in 0.08 s, 13.2 s of speech in 0.10 s, 45 s of speech in 0.29 s; Multilingual 4.5 s of speech in 0.09 to 0.12 s, 13.2 s of speech in 0.18 to 0.20 s, 45 s of speech in 0.87 to 0.92 s; Ultra 4.5 s of speech in 0.08 s, 13.2 s of speech in 0.13 s, 45 s of speech in 0.36 to 0.39 s. Peak app memory 69 to 84 MB. Hot: it had just finished an iOS update and was charging. It is likely faster at rest.

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 12 November 2026.

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.

Benchmark sources: https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2, https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3, https://huggingface.co/moondream/parakeet-ultra, https://github.com/FluidInference/FluidAudio/blob/main/Documentation/Benchmarks.md, https://github.com/FluidInference/FluidAudio/blob/main/Documentation/ASR/ParakeetUltra.md, https://huggingface.co/spaces/hf-audio/open_asr_leaderboard

Source: https://keyspoken.com/models