Released Israeli foods line
Jachnun v1.0
English–Hebrew code-switching
English dictation with Hebrew words dropped in. The Hebrew words come out in romanized Latin letters: makolet, teudat zehut, arnona.
- For English speakers in Israel who mix everyday Hebrew into their English.
- Output is always English text. It never writes Hebrew script.
- About 70% code-switched speech and 30% plain English in training, so plain English does not get worse.
- Earlier proof-of-concept checkpoints were PoC-1 to PoC-5. Jachnun v1.0 is the first named release.
5.42% Code-switched speech stock v2 11.82%
46.9% Unseen Hebrew terms spelled right stock v2 23.8%
1.20% LibriSpeech stock v2 1.23%
Normalised WER on the shared test set, lower is better. Term recall: higher is better.
- Base
nvidia/parakeet-tdt-0.6b-v2- Released
- 2026-10-19
- Training run
r6-en30- Apps
- Hebrish (Android keyboard and desktop apps), Parlance (Android, Linux and Windows)
Released Israeli foods line
Sabich v1.0
Code-switching plus tech vocabulary
The second code-switching model, next to Jachnun. It is trained on Jachnun’s data plus everything else: tech vocabulary, brand names, acronyms, numbers and currency, Irish-English phrasing, disfluencies and self-corrections.
- Named after the Israeli pita with a bit of everything in it.
- Retrained from stock Parakeet, not on top of Jachnun. Every version starts from the same base with the cumulative data.
- 560.5 minutes of my own recordings (373.1 of them code-switching) plus 62.5 minutes of replay. 20 epochs on one Modal L40S, 52 minutes.
- Level with Jachnun on code-switching, a modest gain on tech vocabulary. The extra data did not cost code-switching accuracy, so both models ship in the Hebrish apps.
5.30% Code-switched speech Jachnun 5.42%, stock v2 11.82%
48.1% Unseen Hebrew terms spelled right Jachnun 46.9%, stock v2 23.8%
1.54% Tech dictation Jachnun 1.74%, stock v2 2.51%
1.30% LibriSpeech Jachnun 1.20%, stock v2 1.23%
Normalised WER on the held-out split, lower is better. Term recall: higher is better. On the public benchmark (84 clips) Sabich and Jachnun are within noise: term recall 73.8% vs 73.3%, WER 3.27% vs 3.32%.
- Base
nvidia/parakeet-tdt-0.6b-v2- Released
- 2026-10-20
- Training run
r7-all- Apps
- Hebrish (Android keyboard and desktop apps)
Released Irish foods line
Boxty v1.0
English dictation, no Hebrew
Everyday English dictation without Hebrew: tech vocabulary, brand names, acronyms, numbers and currency in digits, homophones, Irish-English phrasing, and email and message dictation.
- The no-Hebrew sibling of Jachnun. No code-switched, Hebrew-name or religious-term recording went into it.
- Named after Irish foods because the speaker is Irish. Boxty is the Irish potato pancake.
- Best of the set on Hebrew-free English and tech dictation, with no loss on general English.
- On code-switched speech it stays at stock level (11.5%) by design. Use Jachnun for mixed English and Hebrew.
1.04% Hebrew-free English stock v2 1.79%, Jachnun 1.23%
1.35% Tech dictation stock v2 2.51%, Jachnun 1.74%
1.18% LibriSpeech stock v2 1.23%
Normalised WER on the shared test set, lower is better.
- Base
nvidia/parakeet-tdt-0.6b-v2- Released
- 2026-10-20
- Training run
r7-general- Apps
- Parlance (Android, Linux and Windows)
Experimental
Nemotron Streaming (code-switching)
Streaming variant of the code-switching model
The same code-switching task as Jachnun, on NVIDIA’s Nemotron streaming base. Text appears while you speak, in chunks of 80 to 1,120 ms.
- Trained on the same data with the same recipe as Jachnun v1.0, as a side-by-side comparison.
- Slightly less accurate than Jachnun. Use it when you need live, streaming text.
- Streaming costs no measurable accuracy against offline decoding.
5.65% Code-switched speech, streaming 560 ms offline 5.81%
5.8% Code-switched speech, offline Jachnun 5.1%
Normalised WER on the held-out code-switching split, lower is better.
- Base
nvidia/nemotron-speech-streaming-en-0.6b- Released
- 2026-10-19
- Training run
r6-nemotron-en30
Released
Whisper Hebrish Small ACFT
Code-switching for FUTO Keyboard on Android
A phone-sized Whisper fine-tune for English speech with Hebrew words, built to run inside FUTO Keyboard and FUTO Voice Input on Android.
- A single 264 MB whisper.cpp file (Q8_0), the same format as FUTO’s own English model. Import it in FUTO Keyboard settings.
- FUTO’s ACFT procedure was applied after the fine-tune, so it works with FUTO’s short-clip speed-up instead of repeating itself.
- English-only base, so it cannot drift into Hebrew script.
- The phone-sized version of the earlier Whisper Hebrish (Large v3 Turbo) model.
5.8% Code-switched speech, on the phone FUTO stock 11.8%
50% Unseen Hebrew terms spelled right FUTO stock 26%
2.8% Own plain English FUTO stock 4.1%
Normalised WER in the FUTO runtime, lower is better. Term recall: higher is better.
- Base
openai/whisper-small.en- Apps
- FUTO Keyboard (Android)
