Boxty v1.0: English dictation without the Hebrew

Yesterday I released Jachnun, v1.0 of my English-Hebrew code-switching fine-tune of Parakeet. Today there's a sibling. Boxty v1.0 is the same base model, fine-tuned on my voice, but with no Hebrew in it at all.

It's for plain English dictation, and for tech dictation in particular: model and product names, CLI and cloud terms, brand names, acronyms, numbers and currency written in digits, homophones, Irish-English phrasing, and the emails and messages I dictate every day.

Why a version without Hebrew

A lot of what I dictate isn't Hebrish. It's prompts, emails, notes and commit messages, full of model names, product names and command-line terms. The code-switching model handles those reasonably well. But it was trained mostly on mixed speech, so the English and tech data I'd been recording was only ever a supporting act.

Boxty puts that data in front. It was trained on 142.5 minutes of my own phone recordings, filtered so that nothing with a Hebrew word in it got through (no code-switching, no Israeli names, no Yiddish or religious terms). On top of that there's 15.9 minutes of LibriSpeech replay. That's general English read by other people, and it's there so the model doesn't forget what it already knew. Training took 33 minutes on one GPU.

Results

Every model was scored on the same test clips, none of which were trained on. The figure is normalised word error rate (WER), so lower is better.

• Hebrew-free English: Boxty 1.04%, against 1.79% for stock Parakeet and 1.23% for Jachnun.

• Tech dictation: Boxty 1.35%, against 2.51% for stock and 1.74% for Jachnun.

• General English (LibriSpeech): Boxty 1.18%, against 1.23% for stock. Nothing lost.

• Code-switched speech: Boxty stays at stock level, 11.5%. That's on purpose.

A caveat. The Hebrew-free test set is small (57 clips), so I'd treat differences of less than a point as trends rather than results. On tech dictation, Boxty almost halves stock Parakeet's error rate.

Boxty or Jachnun?

The difference is simple. Jachnun knows the code-switching vocabulary: it writes Hebrew words in English sentences in one consistent romanised spelling (makolet, Bituach Leumi, teudat zehut). Boxty doesn't have any of that vocabulary. Give it a Hebrew word and it will write whatever stock Parakeet would have written.

• You mix Hebrew into your English: use Jachnun. Its home is hebrish.app.

• You dictate plain English, especially tech: use Boxty.

Both are tuned to one voice (mine, Irish, living in Israel). Somebody else will see less of a gain.

The names

I've started giving each model line a food code name, because "Parakeet TDT 0.6B v2 fine-tune, run r7-general" isn't something anyone can remember. The code-switching line is named after Israeli foods, so its first release is Jachnun. The no-Hebrew line is named after Irish foods. Boxty is the Irish potato pancake. Later retrains keep the name and bump the version (Boxty v1.1, and so on).

There's also a model trained on everything combined, a kitchen sink. It's now published as Sabich v1.0, the second Hebrish model.

Get it

The model is on Hugging Face. It's gated: access is by request, under the Hebrish terms of use.

Hebrish/Parakeet-TDT-0.6B-v2-Tech-And-Accent

The exact training mix is published as Hebrish/Parakeet-V2-Blend-No-Hebrew, and all the training datasets for these fine-tunes are in one collection.

To use it for dictation, Parlance is my general-purpose dictation app for Android and Linux. Boxty is in its model list from Android 0.5.1 and desktop 0.3.1 (release notes). Existing installs see it at their next update check.

Every fine-tune in this family, Boxty and Jachnun included, is listed on the Carrot Cake AI ASR models page.