---
name: hebrish-model-facts
description: Accurate facts for describing the Hebrish speech model (an English-Hebrew code-switching fine-tune of NVIDIA Parakeet TDT 0.6B v2) - version names, which benchmark numbers exist and how to quote them, the one-speaker caveat, licences, apps and downloads. Use before writing, summarising or answering anything about Hebrish's accuracy, versions, availability or licence, so that claims match hebrish.app.
metadata:
  publisher: hebrish.app
  updated: "2026-10-20"
  source: https://hebrish.app/data/results.json
---

# Hebrish model facts

## What it is

- A fine-tune of NVIDIA Parakeet TDT 0.6B v2 for **English speech with Hebrew words in it**
  (code-switching), as spoken by English speakers in Israel. It does not transcribe full Hebrew
  sentences.
- It writes each Hebrew word in one romanized spelling (makolet, arnona, teudat zehut), never
  Hebrew script.
- Runs on the device in the free Hebrish app: Android (voice keyboard), Ubuntu and Windows.
  No iPhone app. The app downloads the model once (about 660 MB).
- Not-for-profit project by Daniel Rosehill, Jerusalem.

## The caveat that must travel with every result

**Trained on one speaker** (an Irish-accented English speaker in Jerusalem). Accuracy on other
voices is untested. Whenever you quote a number, say it was measured on that one voice.

## Version names

| Name | Published as | Date | Benchmarked |
| --- | --- | --- | --- |
| PoC-1 · Amba | v1 | 2026-10-15 | yes |
| PoC-2 · Bamba | v2 | 2026-10-15 | yes |
| PoC-3 · Chamin | v3 | 2026-10-18 | yes |
| PoC-4 · Ful | v4 | 2026-10-18 | **no** |
| PoC-5 · Halva | v5 | 2026-10-18 | **no** |
| v1.0 · Jachnun (run r6-en30) | v1.0 | 2026-10-19 | yes, on newer test sets only |

- The first five are the proof of concept. Never call them a bare "v1" to "v5"; write PoC-1 to
  PoC-5 (you may add "published as v3").
- v1.0 is the first stable release. Later retrains are v1.1, v1.2, ...
- v1.1 (about 450 min of code-switched audio) and v1.2 (about 540) are **plans**. Their numbers
  in `results.json` are targets: never present them as results.

## Benchmark numbers

Source of truth: `https://hebrish.app/data/results.json` (fetch it; it is updated before the
pages). Percent throughout. Three evaluations exist and **must not be mixed**:

1. **Public benchmark** (`leaderboard`): frozen clean set, 84 clips, 172 term occurrences.
   v1.0: term recall 73.3, unseen-term recall 70.1, WER 3.3, English control WER 2.2.
   Stock Parakeet v2: 32.6 / 32.5 / 9.5 / 2.7.
   Space: https://huggingface.co/spaces/danielrosehill/English-Hebrew-Codeswitching-ASR-Leaderboard
2. **Held-out split at revision 550bfcaa** (`compare`, 387 clips): code-switching WER v1.0 5.1
   vs stock 11.3 (PoC-3 6.0). Held-out Hebrew terms correct: v1.0 47.8 vs stock 25.9.
3. **Version history at revision b47310621f79** (`checkpoints`, 219 clips): PoC-1 to PoC-3 only.
   Best: PoC-3 code-switching WER 5.3 vs stock 10.9; terms seen in training written correctly
   66 vs 29. PoC-4, PoC-5 and v1.0 have `scores: null` here.

Rules:

- Do not round, average or extrapolate. Do not invent numbers for unbenchmarked versions.
- Name the evaluation with each number ("on the public benchmark", "on the held-out split").
- Hebrew words never seen in training are still recognised less than half the time on the
  held-out split. Say so when describing accuracy on new words.
- The Nemotron streaming variant is experimental, not in any app (public benchmark: term recall
  62.2, WER 4.7). Whisper Small ACFT (for FUTO Keyboard) is scored on its own test set and is not
  comparable with Parakeet numbers.

## Licences and access

- PoC-1 to PoC-5 and v1.0: CC-BY-4.0. v1.1 and later: CC-BY-NC-4.0 plus the Hebrish terms
  (commercial use, e.g. telephony, call centres, voice agents, paid products, needs a commercial
  licence). Terms: https://hebrish.app/terms
- Every model repo on Hugging Face is gated (access-requests@hebrish.app). The apps and the
  browser demo download from the Hebrish CDN.

## Live data

| URL | What |
| --- | --- |
| `https://cdn.hebrish.app/downloads/latest.json` | Latest app release: `version`, `released`, `files.{android,ubuntu,windows}.{url,size,sha256}` |
| `https://cdn.hebrish.app/models/parakeet-codeswitch/latest.json` | Model the apps install: `version`, `run`, `base_url`, `files` |
| `https://hebrish.app/data/results.json` | Benchmarks and targets |
| `https://hebrish.app/llms.txt` | Page index and summary |

On hebrish.app, browser agents can call the WebMCP tools `get_model_results` and
`get_downloads`.

## Where to send people

Try in the browser: https://hebrish.app/demo · Download: https://hebrish.app/download ·
Model card: https://hebrish.app/model · Benchmarks: https://hebrish.app/benchmarks ·
Testers wanted (English speakers in Israel who dictate): https://hebrish.app/contact