Hebrish
Technical

Models · k2-fsa · Zipformer2 66M streaming (Hebrew)

Gefilte

Daniel's Hebrew dictation model, for speech that is all Hebrew, in his own Anglo accent. Not part of the Hebrish family: a personal fine-tune and a proof of concept, listed here because it comes from the same work.

Early stage. Gefilte is a proof of concept, not a finished model. PoC-1 was trained on under five minutes of one speaker's recordings, to test whether the pipeline below works at all. Recording continues, and later versions will be trained once there are hours of audio rather than minutes. For Hebrew dictation today, use ivrit.ai's model, which is what the Hebrish apps already offer for Hebrew.

Not part of the Hebrish code-switching models: a Hebrew-only fine-tune of ivrit.ai's streaming Zipformer on Daniel's own non-native Hebrew, read from nikud. PoC-1 was trained to see whether it could learn his surname. It did.

Base modelHebrish/Zipformer2-66M-Streaming-He (k2-fsa)
Current versionGefilte PoC-1, released 2026-11-17
Parameters66M (base model's size)
EvaluationPoC-1 on 8 held-out takes of Daniel's voice: WER 21.2% → 13.5%, his surname recognised in 2 of 2 (base 0 of 2). A sanity check, not a benchmark
Results on the model card
AccessDirect download: no request or sign-in needed. Download on Hugging Face
Downloadevery format and quant (Hugging Face)
LicenceCC-BY-4.0
Used bysherpa-onnx and icefall, by hand
DOIsnone yet
Full model cardresults, recipe, limitations

About it

Why "Gefilte"? It's a joke. Daniel's models are named after food, and an Ashkenazi oleh's Hebrew, read slowly from nikud with the accent intact, could only be gefilte fish.

Hebrish models handle Hebrew words inside English speech. Gefilte is for speech in Hebrew alone. It starts from ivrit.ai's streaming Zipformer, developed by Yakov Kolani and trained on ivrit.ai Hebrew audio, which is already the Hebrew-only model in the Hebrish apps. Gefilte adds Daniel's own recordings on top: Hebrew sentences he reads aloud from a pointed (nikud) version in the recording app, plus the names and dev and AI words he actually needs.

PoC-1 asked one question: can under five minutes of one non-native speaker's recordings teach the model a word it has never heard? The test word was his surname, רוזהיל. The base wrote it as רוזיל and רוזהר. Gefilte PoC-1 got it right in both held-out recordings, and its word error rate on the 8 held-out takes fell from 21.2% to 13.5%. Eight takes from one speaker are a sanity check, not a benchmark, so read the numbers that way. It isn't offered in the Hebrish apps.

How it is made

Six steps, from a written sentence to a model an app can load:

  1. Sentences. Short Hebrew sentences from Tatoeba, graded by difficulty from 1 to 5. On top of these come a priority set (Daniel's surname, and the spoken dev and AI Hebrew he actually uses) and topic sets for particular fields.
  2. Nikud. Each sentence is pointed by Dicta's DictaBERT menaked model, so a non-native reader knows how every word is pronounced, and then checked by script for vav and spelling slips.
  3. A reference reading. A Hebrew text-to-speech voice (Phonikud with its Piper voice) reads each sentence aloud first, so Daniel can hear how it should sound before recording it.
  4. Recording. Daniel reads each sentence from the pointed version in his recording app. The take is stored with the unpointed Hebrew as its transcript, in the public dataset danielrosehill/Gefilte-Hebrew-Sentences (tab takes; the sentences are in tab sentences).
  5. Fine-tuning. The takes train on top of ivrit.ai's streaming Zipformer with the icefall recipe the base was built with. The model starts from ivrit.ai's weights, not from scratch. The newest takes of each priority word, and the dataset's test rows, are held out so the result can be checked on recordings the model has never heard.
  6. Export. The weights are converted to sherpa-onnx int8: the same format, size and streaming setup as the base model, so any app that runs ivrit.ai's model can run Gefilte.

Credit

Gefilte would not exist without ivrit.ai. Its base, ivrit-ai/zipformer2-66m-streaming-he, was developed by Yakov Kolani. It was trained on about 20,000 hours of ivrit.ai's Hebrew audio, labelled with ivrit.ai's own Whisper models; ivrit.ai's open Hebrew speech datasets are on Hugging Face. The base is CC BY 4.0, and so is Gefilte. Hebrish keeps an unchanged copy of it as Hebrish/Zipformer2-66M-Streaming-He, which is what Gefilte is trained from.

Also: Dicta for the nikud model, Phonikud for the reference voice, Tatoeba contributors for the sentences, and k2-fsa for icefall and sherpa-onnx.

Formats

RepositoryFormat and filesVersionForDOI
Hebrish/Gefilte mainicefall checkpoint, sherpa-onnx int8
model.pt 273 MB; sherpa-onnx int8 encoder 70 MB, decoder, joiner · direct download · files
Gefilte PoC-1sherpa-onnx apps, desktop and phonenone yet

Hugging Face issues a DOI per repository, so each format can be cited on its own. Repositories, sizes and DOIs checked on the Hub on 2026-11-17.

Versions

VersionReleasedRunWhereNotes
Gefilte PoC-1 current2026-11-17gefilte-poc1-r2Gefilte84 takes, 4.9 min of Daniel's Hebrew; proof of concept

Access

Open: no access request needed. Download Gefilte from Hugging Face and use it under the CC-BY-4.0 licence.

Download on Hugging Face

Cite

Gefilte has no DOI yet; cite the repository, Hebrish/Gefilte.

@misc{rosehill_gefilte_2026,
  author    = {Daniel Rosehill},
  title     = {Gefilte},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/Hebrish/Gefilte}
}