Model card
The Hebrish model
A fine-tune of NVIDIA Parakeet TDT 0.6B v2 for English with Hebrew words in it. It writes each Hebrew word in one romanized spelling. It was trained on one speaker's voice.
What it does
Many English speakers in Israel drop Hebrew words into their English. Ordinary speech recognition hears those words and guesses an English word that sounds similar. Hebrish learns the Hebrew word and writes it in Latin letters, in the same spelling every time.
Stock ParakeetI want to go to the Micollette today…
HebrishI want to go to the makolet today…
The output is English text. Hebrew words appear romanized, never in Hebrew script, because Parakeet's tokenizer has no Hebrew characters. Casing and punctuation are kept. The spellings follow the Hebrew-Latin-Token-Classifier lexicon.
One speaker's voice. Every training recording is one person: an English speaker in Jerusalem with an Irish accent, reading prompts on an Android phone. It is a proof of concept. How well it works for other voices has not been tested. Help make it multi-speaker.
At a glance
Downloads and formats
All files are on Hugging Face. The main repository holds every format; the separate repositories hold one format each.
| Format | Size | Use it for |
|---|---|---|
| .nemo | 2.47 GB | NeMo, further fine-tuning |
| GGUF Q8_0 recommended | 730 MB | transcribe.cpp v0.1.3 |
| GGUF Q4_K_M | 475 MB | transcribe.cpp, smaller |
| GGUF F16 | 1.2 GB | transcribe.cpp, full precision |
| ONNX fp32 and int8 | — | ONNX Runtime |
| sherpa-onnx int8 | ≈ 661 MB | The Parlance apps |
| MLX fp32 still v3 | — | Apple Silicon |
The MLX build has not been updated since v3; every other format is v5. Quantization check (v5, 2026-10-18): on five held-out test clips, GGUF Q8_0 matched full-precision ONNX on 5 of 5 and Q4_K_M on 4 of 5. That five-clip check is the only on-device accuracy measurement so far. To use it without installing anything, try the demo or the apps.
Results
v4 and v5 not benchmarked The table covers stock Parakeet and versions 1 to 3. Versions 4 and 5 are published but have not been scored yet. The tracker will say when they are.
Scores are on held-out recordings never used in training (test revision b47310621f79, 219 clips, normalised word error rate). Lower WER is better; higher term accuracy is better.
Pick a test above; hover or tab through the points. The hollow points are targets for the next checkpoints, set on 2026-10-19: v6 is planned at 360 minutes of code-switched audio, then a checkpoint every 90 minutes. They are goals, not results. The English tests have a ceiling instead of a goal: no worse than stock. Progress toward the next checkpoint: roadmap.
| Test | Stock v2 | v1 | v2 | v3 |
|---|---|---|---|---|
| Code-switching WER | 10.9% | 6.0% | 5.8% | 5.3% |
| … recorded on phone | 10.7% | 6.0% | 5.6% | 5.1% |
| … desktop mic (30 clips) | 13.8% | 7.1% | 8.6% | 8.6% |
| Hebrew terms seen in training, correct | 29% | 57% | 60% | 66% |
| Hebrew terms never seen, correct | 27% | 47% | 47% | 48% |
| Own English (33 clips) WER | 3.5% | 3.2% | 2.6% | 2.9% |
| LibriSpeech test-clean (200 utts) WER | 1.2% | 1.4% | 1.1% | 1.0% |
| Long-form (9 files, ~3 min) WER | 9.0% | 5.9% | 5.6% | 5.0% |
"Stock v2" is the unmodified NVIDIA base model. "Own English" is plain English in the same speaker's voice, a check that the fine-tune did not damage ordinary English.
Public leaderboard
The leaderboard scores Hebrish and other models on a frozen test set. It is early-stage and a work in progress. This is the clean subset of test set v1: 84 clips, 172 Hebrew term occurrences, 139 distinct terms, 117 of them never seen in Hebrish training.
| Model | Term recall | Unseen-term recall | WER | English control WER |
|---|---|---|---|---|
| Hebrish v3 | 67.4% | 63.2% | 4.0% | 2.5% |
| Hebrish v2 | 65.7% | 60.7% | 4.2% | 2.4% |
| Hebrish v1 | 62.2% | 61.5% | 4.6% | 2.5% |
| Whisper Large V3 | 45.9% | 46.2% | 7.5% | 2.4% |
| Whisper Large V3 Turbo | 40.1% | 39.3% | 8.9% | 2.4% |
| Parakeet TDT 0.6B v2 (stock) | 32.6% | 32.5% | 9.5% | 2.7% |
| Parakeet TDT 0.6B v3 (stock) | 31.4% | 31.6% | 10.5% | 2.7% |
Known issue: 22 of the 139 term labels are English tech words, not Hebrew. That makes term recall 2–6 points higher than a Hebrew-only count would be.
Version history
Each version is the same recipe trained on more recordings. One clip is one read-aloud sentence. Every version is pinned at a fixed revision on Hugging Face. Release notes: changelog.
| v1 | v2 | v3 | v4 | v5 | |
|---|---|---|---|---|---|
| Released | 2026-10-15 | 2026-10-15 | 2026-10-18 | 2026-10-18 | 2026-10-18 |
| Code-switched sentences | 873 | 1,248 | 1,921 | 2,622 | 2,829 |
| Code-switched audio, min | 77.4 | 120.1 | 196.1 | 269.3 | 294.2 |
| Own English, min | 22.3 | 28.7 | 42.9 | 72.5 | 82.6 |
| LibriSpeech replay, min | 11.3 | 17.3 | 28.1 | 38.6 | 42.1 |
| Total clips / min | 1,103 / 111.0 | 1,551 / 166.1 | 2,391 / 267.1 | 3,380 / 380.4 | 3,679 / 418.9 |
| Training time (one L40S) | 19.3 min | 23.0 min | 28.9 min | 33.1 min | 38.0 min |
| Benchmarked | yes | yes | yes | no | no |
| Pinned revision | ebc0e7aff1 |
e1541065f8 |
728e6f88d4 |
2e02382e7d |
5522d9e86e |
- v1
- First release, 77.4 minutes of code-switched audio. Code-switching WER fell from 10.9% (stock) to 6.0%.
- v2
- More recordings (77.4 → 120.1 min, 1,248 clips). Code-switching WER 5.8%.
- v3
- 76 more minutes. Code-switching WER 5.8% → 5.3%, long-form WER 5.6% → 5.0%. The last benchmarked version.
- v4
- 73 more minutes (196 → 269 min, 2,622 clips). The training mix reached its 70/20/10 target for the first time. Not benchmarked.
- v5
- 25 more minutes (269 → 294 min, 2,829 clips) and 10 more minutes of the speaker's own English (73 → 83 min). A smaller step than the usual 1.5×, released by choice. Not benchmarked.
Each version's validation WER was measured on a different validation set, so those numbers are not comparable across versions and are not listed here.
Training recipe
v5 is a full fine-tune of the stock base model. All five versions share the core settings: 20 epochs, learning rate 3e-5, batch size 16, seed 7, and a 30% English share target.
- NeMo 2.7.3, AdamW, cosine schedule with 10% warmup, bf16.
- One L40S GPU; v5 trained in 38 minutes. Each run cost under a dollar.
- Best checkpoint chosen by validation WER (v5: epoch 16).
- v5 mix by audio duration: 70.2% code-switching, 19.7% the speaker's own English, 10.1% LibriSpeech replay (target 70/20/10).
- v5 code-switched audio: 294 minutes in 2,829 clips. 2,390 clips (267.4 min) from the phone capture app, 439 clips (26.8 min) imported from desktop-mic recordings. Plus 83 minutes of own English (652 clips) and 42 minutes of LibriSpeech replay.
- Evaluation set at v5's revision: 311 code-switching clips, term-disjoint, 265 held-out terms.
Why mix in plain English? Fine-tuning on code-switched speech alone tends to damage a model's ordinary English. The speaker's own English and LibriSpeech replay guard against that. By v3, the English scores were as good as or better than stock.
Intended use
Dictation by English speakers who use Hebrew words, running on your own device. It is a proof of concept and a single-speaker model. It works best for voices like the speaker's. If yours is very different, stock Parakeet may do better; the Parlance apps let you switch.
Limitations
- One speaker. Gains on other voices are untested.
- Hebrew words it never saw in training are recognised less than half the time (measured at v3).
- Spellings follow one convention, the Hebrew-Latin-Token-Classifier lexicon. Your preferred spelling may differ.
- On-device accuracy has not been measured beyond the five-clip GGUF check.
- v4 and v5 are not benchmarked.
Roadmap
- v6 at 360 minutes of code-switched audio. Recordings are being collected now.
- v7 at 450 minutes, then a new version every 90 more minutes.
- Retrains go in steps of about 1.5× the data.
- More speakers. That needs your recordings.
Also: a streaming variant on Nemotron
experimental not benchmarked not in any app
There is a second fine-tune on nvidia/nemotron-speech-streaming-en-0.6b, a 0.6B cache-aware FastConformer with an RNNT decoder. Its base model can stream, transcribing in chunks of 80 to 1,120 ms while you speak.
- Same data and recipe as Parakeet v5. Trained 2026-10-18, 34.6 minutes on an L40S, best epoch 15 of 20.
- Validation WER 6.4% (raw, offline decoding), against 4.9% for the Parakeet v5 run on the same validation set.
- It was trained offline with full context, not streaming-aware. Streaming accuracy has not been measured.
- Only a .nemo file is published; no GGUF or ONNX yet.
- Different licence: the NVIDIA Open Model License Agreement, not CC-BY-4.0. Not made or endorsed by NVIDIA.
Nemotron variant on Hugging Face