Version 2 of Whisper Hebrish is out. It's a fine-tune of openai/whisper-large-v3-turbo for English speech with Hebrew words dropped in, the way English speakers in Israel actually talk ("I need to renew my teudat zehut before we go to the makolet"). The output is English, and each Hebrew word comes out in one canonical Latin spelling. It doesn't transcribe Hebrew itself.
v1 went up in November 2025 and was trained on about 25 minutes of desktop recordings. v2 is the Whisper sibling of my Parakeet code-switching fine-tune, trained on the same data.
What changed
Data: 1,929 clips (215 minutes) from Read-Aloud-Voice-Sentences, mostly recorded on my phone, against 516 desktop clips for v1.
Spelling: one canonical romanisation per term, where v1's spellings were ad hoc.
English: 29% of the training mix is plain English, so the model doesn't forget how to transcribe it.
Test split: term-disjoint, so some of the Hebrew terms in the test set never appear in training.
Results
Term-disjoint test set, greedy decoding. Stock turbo → v1 → v2:
Code-switching WER: 14.7% → 12.1% → 6.7%
Unseen Hebrew terms transcribed correctly: 36.7% → 48.6% → 53.2%
Hebrew terms seen in training, transcribed correctly: 43.8% → 50.0% → 76.8%
Long-form dictation WER (normalised): 15.4% → 20.7% → 9.2%
My own plain English (normalised): 4.0% → 4.1% → 2.9%
The one place it gives ground is LibriSpeech (1.8% stock against 2.2%), a fair trade for a personal dictation model. v1 is still available at the v1 tag on the repo.
Get it
The repo has Transformers weights plus whisper.cpp builds (f16, q8_0 and q5_0). On ten test clips all three gave output identical to the PyTorch model. q8_0 (874 MB) is the one to use. Always force English, or Whisper may switch to Hebrew script:
whisper-cli -m ggml-whisper-hebrish-v2-large-v3-turbo-q8_0.bin -l en -f clip.wav
It was trained on one speaker (me), so other accents are untested. For a phone-sized version that runs inside FUTO Keyboard, see Whisper Hebrish Small ACFT. The Parakeet equivalent is covered in v3 of the Parakeet code-switching fine-tune.
