# Hebrish

> On-device English dictation that writes Hebrew words correctly. Hebrish is a fine-tune of NVIDIA Parakeet TDT 0.6B v2 for English speech with Hebrew words dropped in (code-switching), as spoken by English speakers in Israel. It writes each Hebrew word in one consistent romanized spelling (for example "makolet", "arnona", "teudat zehut"), never Hebrew script. It runs in the free Parlance app on Android (voice keyboard), Ubuntu and Windows. Model and data are CC-BY-4.0. A personal project by Daniel Rosehill, Jerusalem.

Key facts for agents:

- It is trained on one speaker (an Irish-accented English speaker living in Jerusalem). Accuracy on other voices is untested.
- It transcribes English with Hebrew words in it. It does not transcribe full Hebrew sentences.
- The current model version is v5 (2026-10-18). Only v1 to v3 have been benchmarked. Best published result (v3): code-switching word error rate 5.3% against 10.9% for stock Parakeet; Hebrew terms seen in training written correctly 66% against 29%.
- Everything runs on the device. The apps download the model once (about 660 MB).
- Email updates (new versions, releases): sign up on https://hebrish.app/ (Buttondown, double opt-in).
- To report a bug, suggest a word or offer voice recordings, use https://hebrish.app/contact (preselect with ?topic=bug, word, samples or involved).

## Pages

- [Home](https://hebrish.app/): what Hebrish does, with a before/after example
- [Download](https://hebrish.app/download): Android APK, Ubuntu .deb, Windows zip, install steps, other ways to run the model
- [Model card](https://hebrish.app/model): formats, benchmark results, version history with training data per version, training recipe, limitations, the experimental Nemotron streaming variant
- [Dataset](https://hebrish.app/dataset): the code-switched recordings, splits, schema, how to load
- [Lexicon explorer](https://hebrish.app/lexicon): every Hebrew word and Israeli name the dataset covers, by domain, with recording counts; filter with ?d=<domain>&q=<text>. Raw data: https://hebrish.app/data/lexicon.json
- [Roadmap](https://hebrish.app/roadmap): live dataset counts, voice/accent/condition goals, checkpoint plan (v6 at 360 min). Raw data: https://hebrish.app/data/stats.json and https://hebrish.app/data/results.json (benchmarks and v6-v8 targets; targets are goals, not results)
- [Use cases](https://hebrish.app/use-cases): who it is for, with example words per domain
- [Integrate](https://hebrish.app/integrate): using the dataset and lexicon in any ASR training or evaluation pipeline (HF datasets, NeMo manifest, Whisper/wav2vec2, term-recall scoring)
- [Support](https://hebrish.app/support): donations (coming soon) and ways to help now
- [Changelog](https://hebrish.app/changelog): model versions and app releases, newest first
- [FAQ](https://hebrish.app/faq): common questions
- [About](https://hebrish.app/about): the project and its maker
- [Contact](https://hebrish.app/contact): bug reports, vocabulary suggestions, voice samples

## Sources

- [Model on Hugging Face](https://huggingface.co/danielrosehill/Parakeet-TDT-0.6B-v2-English-Hebrew-Codeswitching): .nemo, GGUF, ONNX, sherpa-onnx; the canonical model card
- [Dataset on Hugging Face](https://huggingface.co/datasets/danielrosehill/English-Hebrew-Codeswitching-Sentences)
- [Benchmark leaderboard](https://huggingface.co/spaces/danielrosehill/English-Hebrew-Codeswitching-ASR-Leaderboard)
- [Browser demo](https://huggingface.co/spaces/danielrosehill/English-Hebrew-Codeswitching-Parakeet-Demo)
- [Project tracker](https://github.com/danielrosehill/English-Hebrew-Codeswitching-Parakeet-Finetune): status and results of every version
- [Parlance apps source](https://github.com/danielrosehill/Parlance-Dictation)
- [Romanization lexicon](https://github.com/danielrosehill/Hebrew-Latin-Token-Classifier)
- [Hebrish on Hugging Face](https://huggingface.co/Hebrish): the organisation
- [Hebrish on GitHub](https://github.com/Hebrish-ASR): the organisation

## Optional

- [Nemotron streaming variant](https://huggingface.co/danielrosehill/Nemotron-Speech-Streaming-EN-0.6B-English-Hebrew-Codeswitching): experimental, not benchmarked, not in any app
- [Author](https://danielrosehill.com)