AI · open models
Small, open models that run on your own device. Most are about one problem: speech recognition that gets the Hebrew words right when English speakers in Israel mix them in. Everything is on Hugging Face, with results you can check.
Series · English–Hebrew
Speech and text models for English with Hebrew mixed in: the way English speakers in Israel actually talk. Speech fine-tunes on open models from NVIDIA, Qwen, FUTO, OpenAI and Meta, from a 28M-parameter wearable model to a 2.5B desktop one, plus a text model that finds Hebrew written in Latin letters. The speech models are named after Israeli food.
Yes, you can pay with the Cartisa Shrai, but if you pay him as Uman, there's a small discount.
Yes, you can pay with a kartis ashrai, but if you pay mezuman there's a small discount.
Pick up the forms for the Rishayoniski from Chana and ask her about the Vad and Hershikun while you're there.
Pick up the forms for the rishayon iski from Chana, and ask her about the vaad in her shikun while you're there.
Hi Avi, I checked with the Sevet and the answer is yes. We can have the demo ready for the Reicheva client by Thursday.
Hi Avi, I checked with the tzevet and the answer is yes, we can have the demo ready for the Be'er Sheva client by Thursday.
S. D. Rabinowitz told me I still need to fill out a TOFS for the tax authority before the end of the month.
Esti Rabinowitz told me I still need to fill out a tofes for the tax authority before the end of the month.
Did you see that the Mem Shalah just approved another round of funding for the cyber programme? Let's read the details tonight.
Did you see that the memshala just approved another round of funding for the cyber program? Let's read the details tonight.
My Makshev died again this morning, so I'll be working for my phone until I get it to the shop.
My machshev died again this morning, so I'll be working from my phone until I get it to the shop.
Verbatim transcripts from the public benchmark, round of 2026-10-19. More on hebrish.app.
The recommended model in the Hebrish apps. Trained on everything: the code-switched speech plus tech vocabulary, brand names, numbers and other task recordings.
73.8% term recall, 3.3% WER
Trained on English–Hebrew Code-Switching SentencesAll training datasets (collection)
NVIDIA Parakeet TDT 0.6B v2 · 0.6BCard page →
The pure code-switching line: English with Hebrew words. In the apps next to Sabich; every on-device format.
73.3% term recall, 3.3% WER
Trained on English–Hebrew Code-Switching SentencesRead-Aloud Voice Sentences
NVIDIA Parakeet TDT 0.6B v2 · 0.6BCard page →
Sabich's training data on Qwen3-ASR 1.7B, an ASR model with an LLM decoder. The most accurate Sabich, for computers with a GPU. Not in the Hebrish apps yet.
5.4% code-switching WER (stock Qwen3-ASR 10.0%)
Trained on Read-Aloud Voice Sentences (revision e918e381, blend all)All training datasets (collection)
Qwen Qwen3-ASR 1.7B · 2.0BCard page →
Sabich's training data on NVIDIA Canary-Qwen 2.5B, a speech encoder feeding Qwen3-1.7B. For desktop GPUs: the strongest of the Sabich variants on plain English, behind Sabich-Qwen3-ASR-1.7B on code-switching. Not in the Hebrish apps yet.
6.9% code-switching WER (Sabich-Qwen3-ASR-1.7B 5.4%), 2.4% on the speaker's own English
Trained on Sabich training blend
NVIDIA Canary-Qwen 2.5B · 2.56Bhebrish.app ↗
All of Sabich's training on NVIDIA's smallest Parakeet, about a fifth of the size. For low-end phones, watches and wearables. Not in the Hebrish apps yet.
Not benchmarked yet
Trained on Sabich training blend
NVIDIA Parakeet TDT-CTC 110M · 114Mhebrish.app ↗
Sabich's training data on the tiny asr4all streaming models, 28M and 56M parameters, on the CPU. For wearables and very low-power boards. Numbers come out as words. Not in the Hebrish apps yet.
9.1% (-M) and 10.1% (-S) code-switching WER (stock asr4all-m 19.5%)
Trained on Sabich training blend
FUTO asr4all S and M · 28.2Mhebrish.app ↗
The streaming line: text appears while you speak. Shakshuka v1.0 is trained on Sabich's data (code-switching plus tech) on NVIDIA's cache-aware streaming model.
v1.0 not benchmarked yet; PoC-2: 62.2% term recall, 4.7% WER (offline)
Trained on Read-Aloud Voice Sentences (revision e918e381, blend all)All training datasets (collection)
NVIDIA Nemotron Speech Streaming EN 0.6B · 0.6BCard page →
A phone-sized Whisper (244M) with FUTO's ACFT, for people who already use FUTO Keyboard on Android.
5.8% code-switching WER on its own test set (FUTO runtime)
OpenAI Whisper small.en · 242Mhebrish.app ↗
Full fine-tune of Whisper large-v3-turbo (809M). Where Hebrish started, in November 2025; the larger sibling of the Small ACFT model.
Version 1: 54.1% term recall, 6.5% WER
OpenAI Whisper large-v3-turbo · 809Mhebrish.app ↗
Text-to-speech for code-switched English: Irish-accented English that pronounces the Hebrew words in it correctly. A full fine-tune of Qwen3-TTS; the first checkpoint, Meorav PoC-1, is an experiment.
No benchmark yet; listening verdict pending
Qwen Qwen3-TTS 12Hz 1.7B Base · 1.9Bhebrish.app ↗
Marks each romanized Hebrew word in an English sentence, so text-to-speech can send it to a Hebrew voice. Span F1 0.941 on terms it never saw in training.
Span F1 0.941 on terms it never saw in training
Trained on Hebrish/hebrew-latin-code-switching (4,737 sentences)
Meta XLM-RoBERTa base · 277MCard page →
Every Hebrish model, dataset and demo → Download the apps on hebrish.app ↗
The public English–Hebrew code-switching benchmark: the share of Hebrew terms each model writes with the canonical romanized spelling (test set v1, clean subset: 84 clips, 172 Hebrew term occurrences). A model can score a low word error rate and still mangle every Hebrew word, so this is the headline number.
| Model | Hebrew terms right | WER |
|---|---|---|
| Hebrish Sabich v1.0 | 73.8% | 3.3% |
| Hebrish Jachnun v1.0 | 73.3% | 3.3% |
| Hebrish Shakshuka PoC-2 (Nemotron streaming) | 62.2% | 4.7% |
| Whisper Hebrish v1 (known training-data overlap) | 54.1% | 6.5% |
| Whisper Large V3 | 45.9% | 7.5% |
| Whisper Large V3 Turbo | 40.1% | 8.9% |
| Parakeet TDT 0.6B v2 | 32.6% | 9.5% |
| Parakeet TDT 0.6B v3 | 31.4% | 10.5% |
All test clips are one speaker reading aloud, so treat gaps of a few points as noise. Leaderboard as of 21 October 2026: Hugging Face Space · benchmark code · test set · all my benchmarks.
The models are parts of two pipelines, both meant to run on-device.
Series · dictation toolkit
Models for the rest of a voice-note pipeline: plain English dictation in my voice, then small text models that tidy and route the transcript. Built to run on-device after the speech model.
English and tech dictation without Hebrew: model names, CLI terms, brands, numbers in digits.
2.51% → 1.35% tech dictation WER
Trained on Parakeet V2 Blend: No Hebrew
NVIDIA Parakeet TDT 0.6B v2 · 0.6BCard page →
Adds paragraph breaks to flat ASR output and changes nothing else.
0.78 break F1 (untrained: 0.00)
Trained on danielrosehill/ASR-Paragraph-Breaks
Google Gemma 3 270M · 270MCard page →
Puts the key parts of a dictated email in bold and changes nothing else. Runs after Paragraph Breaks.
0.62 word F1 (untrained: all fallbacks)
Trained on danielrosehill/Email-Key-Line-Bolding
Google Gemma 3 270M · 270MCard page →
Reads a transcript and predicts its intent (shopping list, to-do, email, AI prompt) so an agent can route it.
0.56 micro F1, placeholder release
Trained on Voice Notes
BAAI bge-base-en-v1.5 · 110MHugging Face ↗
Series · text privacy
Small models that make real text safer to share. They rewrite it rather than redact it, so it still reads naturally.
Swaps names, places, dates and contacts for close, plausible substitutes, the same one every time. Not safe to run unsupervised yet.
9.3% of identifying details left unchanged (untrained: 85.6%)
Alibaba Qwen3-0.6B · 0.6BCard page →
From my Hugging Face profile, most recently updated first, including each model's export formats. This list refreshes daily, so a brand-new model can take a day to appear. Hebrish releases live on the Hebrish organisation; the Hugging Face index lists both accounts, with datasets, Spaces and collections too.
Work with me
Everything on this site is open source and built in public. If you want the same kind of work done for your organisation, I take on consulting and build projects through DSR Holdings.