Hebrish: English-Hebrew code-switching
Collection
Every Hebrish model, dataset and Space, including those still on danielrosehill. hebrish.app • 34 items • Updated
You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access to this model is by request and is governed by the Hebrish terms of use (https://hebrish.app/terms). Non-commercial use is free; commercial use needs a commercial licence from access-requests@hebrish.app.
Log in or Sign Up to review the conditions and access this model content.
Sabich v1.0: Parakeet TDT 0.6B v2 for English-Hebrew code-switching, tech and general dictation
A personal speech-recognition fine-tune trained on everything: English with Hebrew words in it (code-switching: teudat zehut, arnona, makolet), tech vocabulary (model and product names, CLI and cloud terms), brand names, acronyms, numbers and currency in digits, Irish-English phrasing and everyday email and message dictation, all in one speaker's voice. It is the kitchen-sink sibling of Jachnun v1.0, the pure code-switching model, and of Boxty v1.0, the no-Hebrew model.
Name. Each fine-tune line has a code name from a food, so models are easy to tell apart: the English-Hebrew code-switching line is named after Israeli foods (v1.0 is Jachnun), the no-Hebrew line after Irish foods (Boxty), and this kitchen-sink line is Sabich, the Israeli pita stuffed with a bit of everything. Later retrains of this line keep the name and bump the version (Sabich v1.1, …); the training run id is r7-all.
A full fine-tune of nvidia/parakeet-tdt-0.6b-v2 on one
speaker's read-aloud phone recordings (an Irish English speaker living in Israel), run r7-all, trained
2026-10-20. Output keeps v2's casing and punctuation; numbers are written in digits.
Sibling models from the same data revision: Jachnun v1.0, English-Hebrew code-switching · Boxty v1.0, English without Hebrew.
Access and terms of use. Governed by the Hebrish terms of use; access is by request (button on this page, or access-requests@hebrish.app). Non-commercial use is free; commercial use needs a commercial licence.
Results
All models scored on the same test clips (dataset revision f956dbb184c2, blend all test sets;
never trained on, never used for checkpoint selection). WER is normalised (lowercase, no punctuation).
| Model | Hebrew-free English WER (57 clips) | tech WER | own English WER (84) | LibriSpeech WER (200) | code-switching WER (422) | held-out Hebrew term recall | long-form WER |
|---|---|---|---|---|---|---|---|
stock nvidia/parakeet-tdt-0.6b-v2 |
1.79% | 2.51% | 2.81% | 1.23% | 11.82% | 23.77% | 10.34% |
Jachnun v1.0, code-switching (r6-en30) |
1.23% | 1.74% | 1.56% | 1.20% | 5.42% | 46.91% | 5.84% |
Boxty v1.0, no Hebrew (r7-general) |
1.04% | 1.35% | 2.56% | 1.18% | 11.52% | 23.46% | 11.20% |
this model (r7-all) |
1.23% | 1.54% | 1.62% | 1.30% | 5.30% | 48.15% | 5.77% |
This model has the lowest code-switching WER and the highest held-out Hebrew term recall of the set, and is close to Boxty on tech dictation. LibriSpeech is 0.07 points above stock (about 3 words in 200 clips), within noise but a strict fail of the forgetting rule; general English is otherwise not lost. Differences from Jachnun v1.0 on code-switching are small (within about one WER point). Pick it if you dictate both mixed English-Hebrew and tech-heavy English.
Public benchmark (leaderboard, frozen test set v1, clean subset: 84 clips, 172 Hebrew term occurrences; scored 2026-10-20 on the weights in this repo, first published at revision 640c1483 of its earlier name): term recall 73.8% (Jachnun v1.0: 73.3%), unseen-term recall 68.4% (70.1%), WER 3.27% (3.32%), English control WER 1.99% (2.17%). Level with Jachnun v1.0 within noise. None of the benchmark clips are in this model's training or validation data.
In the Hebrish apps it is the model called Sabich in the model picker; the app build is Hebrish/Sabich-sherpa-onnx-int8 (sherpa-onnx int8, open, no access request).
Training data
The exact training mix is published as Hebrish/Parakeet-V2-Blend-All (tag r7-all).
- Every one of the speaker's takes, code-switching and English: code-switching sentences, Hebrew names, Yiddish and religious terms, and every task subset (tech, brand names, acronyms, expanded acronyms, Irish English, currency, numerics, hard sentences, disfluency, repeated words, singing, weird voices, whispered, web addresses, spelled-out) except
name-cipher, whose targets contradict the rest on purpose. No separate own-English pool: the English is the speaker's own Hebrew-free takes inside the primary set. - Mix: 560.5 min of the speaker's takes (373.1 min code-switching) plus 62.5 min of LibriSpeech-PC replay (cased, punctuated; CC-BY-4.0) = 623.0 min, 5,334 clips.
- Validation: 1 in 16 of the speaker's train takes by id hash, used only to pick the checkpoint.
Training
NeMo 2.7.3 on one L40S (Modal), bf16. AdamW, lr 3e-05 (cosine, 10% warmup), 20 epochs,
batch 16, seed 7. Best checkpoint best-epoch=19-val_wer=0.0463.ckpt,
52.1 min of training.
Usage
import nemo.collections.asr as nemo_asr
from huggingface_hub import hf_hub_download
path = hf_hub_download("Hebrish/Sabich", "Sabich.nemo")
model = nemo_asr.models.ASRModel.restore_from(path)
print(model.transcribe(["clip.wav"])[0].text) # 16 kHz mono
GGUF builds for transcribe.cpp (Offline Voice Input on Android, Handy on Linux) are in Hebrish/Sabich-GGUF: Q8_0 (recommended),
Q4_K_M and F16.
Limitations
- One speaker. It is adapted to one voice, accent and phone; others will see less benefit.
- The test sets are small (Hebrew-free English is 57 clips), so differences under about one WER point are within noise.
- Disfluent and sung takes were trained on with clean targets, so the model may drop some fillers and repeated words.
- LibriSpeech WER is 0.07 points above stock (1.30% vs 1.23%): noise-level, but not an improvement on general read English.
- Its best epoch was the last of 20; a 30-epoch run (
r7-all-e30) exists but is not evaluated or published. - Numbers on the test set are 8 tokens only; the number score is not meaningful yet.
Licence
CC-BY-4.0, from the base model's licence, with use governed by the Hebrish terms above. Base model by NVIDIA.
- Downloads last month
- -