FAQ

Voice dictation for English speakers in Israel. You speak English with Hebrew words in it, and it writes the Hebrew words correctly, in Latin letters: makolet, not Micollette.

It is a fine-tune of NVIDIA's Parakeet TDT 0.6B v2 speech model. It runs on your own phone or computer, inside the free Parlance app. You can also try it in the browser first.

It's the mix of English and Hebrew that English speakers in Israel actually use. You speak English, and drop in Hebrew words where they fit: arnona, teudat zehut, kupat holim.

Linguists call switching languages mid-sentence code-switching. That's the technical name for what this model handles.

No. Hebrish is an English model that knows Hebrew words. It expects English sentences with Hebrew words in them. Speak a full sentence of Hebrew and you will get English-looking guesses, the same as with any English model.

For Hebrew-only dictation, use a Hebrew speech model.

Maybe, and nobody knows yet. All of the training recordings are one speaker's: an Irish-accented English speaker living in Jerusalem. It works best for voices like that. How it does on other voices has not been tested.

If it does worse than stock Parakeet for you, switch back. If you'd like it to work for voices like yours, send some recordings.

Through the Parlance app: an Android voice keyboard that works with HeliBoard, and tray apps for Ubuntu and Windows. The Ubuntu build is tested on Ubuntu 26.04 (amd64). The Windows build is early and not yet tried on real hardware. Links are on the download page.

On a Mac with Apple Silicon there is an MLX build, which is still v3. Any other app that loads a Parakeet GGUF file can use the GGUF builds, for example Offline Voice Input on Android or Handy on Linux.

Android 9 or newer, on an arm64 phone.

Parlance downloads the model it needs (about 660 MB). If you use another app with the GGUF files instead, take the Q8_0 file (730 MB) unless your phone has 4 GB of RAM or less; then take Q4_K_M (475 MB). Speed on phones has not been measured yet.

In Parlance, open the model picker and choose stock Parakeet instead of the English-Hebrew code-switching model. It's one tap, and you can switch back the same way.

Speech models write the words they were trained on. An English model has never heard makolet, so it writes the English-looking words that sound closest. Stock Parakeet wrote "Micollette".

Hebrish was trained on recordings of English sentences with Hebrew words in them, so it has heard the Hebrew words and learned to write them.

Parakeet v2's vocabulary has no Hebrew characters, so the model can't write Hebrew script. It writes each Hebrew word in Latin letters instead.

That also keeps a sentence in one direction. Mixing right-to-left Hebrew into left-to-right English often scrambles word order in messages and documents.

One consistent spelling per word, taken from a public lexicon of 1,164 terms. Every training transcript follows it, so the model always writes, for example, kupat holim, not a different spelling each time.

If you prefer another romanization, the model will still tend to write the lexicon's one.

The latest measured release is v3. On held-out recordings it was never trained on:

• Word error rate on English with Hebrew words: 10.9% for stock Parakeet, 5.3% for Hebrish v3.
• Hebrew words seen in training, written correctly: 29% stock, 66% v3.
• Hebrew words never seen in training, written correctly: 27% stock, 48% v3.
• Plain English from the same speaker, word error rate: 3.5% stock, 2.9% v3.

Versions 4 and 5 are published but not yet benchmarked. All of these numbers are for one speaker's voice. Full results are on the model page.

Mostly the amount of training audio. Each version is trained from stock Parakeet on more recordings: v3 on 196 minutes of code-switched speech, v5 on 294 minutes. The current release is v5.

What changed in each release is on the changelog. How each one was trained is on the model page.

There is an experimental one, built on NVIDIA's Nemotron Speech Streaming model with the same data and recipe as v5. It is not benchmarked, it is published only as a .nemo file, and no app uses it yet.

Transcription runs entirely on your device. Your audio is not uploaded to transcribe it.

The app needs the internet once, to download the model (about 660 MB). After that it checks now and then for a newer version of the model.

Yes to both. The apps are free, and their source code is public. The model is released under CC-BY-4.0, the same licence as the NVIDIA base model, and so is the training dataset.

Use the bug report form. Say what you said, what it wrote, and which app and platform you used.

Yes. Suggest a word, with what it means and how you'd use it in a sentence. A word that gets added gets one spelling in the lexicon, like every other term.

Yes, and this helps more than anything else. The model has only ever heard one voice. Recordings from other speakers are the only way to find out how it does on other voices, and to train it beyond one.

Get in touch about recording some sentences.

Daniel Rosehill, in Jerusalem, as a personal project. It is not made or endorsed by NVIDIA. More on the about page.