Model · speech-to-text
Hebrish (Sabich v1.0): Parakeet TDT 0.6B v2 for code-switching plus tech dictation
The recommended Hebrish model: everything Jachnun was trained on, plus tech vocabulary, brand names, acronyms, numbers and currency, Irish-English phrasing and the rest of my task recordings.
Which model? Sabich and Jachnun are level on code-switching; Sabich is slightly better on tech vocabulary, which is why it is the recommended pick. Both run in the Hebrish apps, so switch in the model list if Jachnun suits you better. See Jachnun →
About the model
Sabich is the second Hebrish model: a fine-tune of NVIDIA Parakeet TDT 0.6B v2 trained on Jachnun's data plus everything else I have recorded. That means the code-switched speech, plus tech vocabulary, brand names, acronyms, numbers and currency, Irish-English phrasing, disfluencies and self-corrections.
It is not trained on top of Jachnun. Every model in the family is retrained from stock Parakeet v2 on the cumulative data: PoC-1 to PoC-5, then Jachnun v1.0, then Sabich v1.0. Boxty, a separate fine-tune for English without Hebrew (not part of Hebrish), starts from the same base.
The question was whether adding all the other data would cost code-switching accuracy. It did not. On the public benchmark Sabich and Jachnun are level within noise, and Sabich is slightly better on tech vocabulary. That settled shipping both models in the Hebrish apps.
Sabich is named after the Israeli pita with a bit of everything in it. The code-switching line uses Israeli food names.
The app: Hebrish
The Hebrish apps (Android voice keyboard, tray apps for Ubuntu and Windows) recommend Sabich, with Jachnun and stock Parakeet in the same model list. They download the model once and transcribe fully offline. Source and every release: github.com/Hebrish-ASR/Hebrish-App.
Linux (Debian / Ubuntu) about 11 MB + 660 MB model
Install with sudo apt install ./hebrish_*_amd64.deb (the file you downloaded). Ubuntu 26.04, amd64.
Download .deb, latest version →Android about 15 MB + 660 MB model
arm64 phones, Android 9+. Works as the voice input behind HeliBoard's mic key, or try it straight away in the app's own notepad. Also on the Hebrish F-Droid repository, which keeps it updated.
Download APK, latest version →Windows about 69 MB + 660 MB model
Unzip and run Hebrish.exe. Not yet tested on real hardware, and unsigned, so SmartScreen may warn.
Download .zip, latest version →Results
v1.0 · Sabich, run r7-all, evaluated 20 October 2026. Lower WER is better; a higher share of terms correct is better.
| Test set | Stock v2 | Fine-tune v1.0 · Sabich |
|---|---|---|
| Code-switching, held-out split (normalised WER) | 11.82% | 5.30% |
| Hebrew terms held out of training: share correct | 23.8% | 48.1% |
| Tech dictation (normalised WER) | 2.51% | 1.54% |
| LibriSpeech (normalised WER) | 1.23% | 1.30% |
- Against Jachnun v1.0 on the same held-out split (same dataset revision): code-switching WER 5.30% against 5.42%; held-out terms 48.1% against 46.9%; tech WER 1.54% against 1.74%; LibriSpeech 1.30% against 1.20%.
- Public benchmark (84 clean clips, round 4), Sabich against Jachnun: Hebrew term recall 73.8% against 73.3%; unseen-term recall 68.4% against 70.1%; WER 3.27% against 3.32%; English control WER 1.99% against 2.17%.
- Read: level on code-switching (the differences are noise), a modest gain on tech vocabulary, and no loss of code-switching accuracy from the extra data.
- WER is normalised: lowercase, punctuation removed.
Training
- v1.0 · Sabich: 560.5 min of my own recordings (373.1 min of them code-switching) plus 62.5 min of LibriSpeech replay. 20 epochs on one Modal L40S, 52 minutes. Run name r7-all.
- Full fine-tune from stock Parakeet TDT 0.6B v2, not continued training from Jachnun. Jachnun v1.0 (r6-en30) used 339.2 min of code-switching, 95.3 min of my own English and 48.5 min of LibriSpeech replay.
Training data
Checkpoints and roadmap
Where the model is now and the checkpoints still to come. Planned stages are the current plan, not promises.
- v1.0 · Sabich (r7-all) Done 20 October 2026
First release: Jachnun's data plus every other task recording, 560.5 min of own recordings plus 62.5 min replay, trained from stock v2. Added to the Hebrish apps' model list.
- Recommended in the apps Done 21 October 2026
Sabich became the recommended pick in the Hebrish apps, with Jachnun kept alongside it.
- Sabich on other bases Done 21 October 2026
The same training data on four more bases: Qwen3-ASR 1.7B (5.4% code-switching WER, stock 10.0%) and Canary-Qwen 2.5B (6.9%) for desktop GPUs, Parakeet TDT-CTC 110M for small devices, and asr4all S and M (10.1% and 9.1%) for wearables. Not in the apps yet. Shakshuka v1.0, the streaming line, is trained on it too.
Limitations
- Trained on one voice so far: an English speaker with an Irish accent living in Jerusalem. Recordings from more speakers are the next step; contribute yours at hebrish.app/contact?topic=samples.
- On code-switching it is level with Jachnun, not better: Hebrew terms the model never heard in training are still right only about half the time.
- The model repos are gated, with access by request.
Related
- Sabich on hebrish.appWhat it is, how it compares with Jachnun, and the files
- Sabich-Qwen3-ASR-1.7BThe most accurate Sabich, for computers with a GPU
- Sabich-Canary-Qwen-2.5BDesktop GPU; the strongest Sabich on plain English
- Sabich-Parakeet-110M and Sabich-asr4allSmall-device and wearable builds of the same data