Hebrish: English-Hebrew code-switching Parakeet
Two fine-tunes of nvidia/parakeet-tdt-0.6b-v2
for English speech with Hebrew words dropped in, the way English speakers in Israel talk:
"pick up bamba from the makolet", "I have an appointment at Misrad Hapnim". Hebrew words come
out romanized (v2's tokenizer has no Hebrew letters), with v2's casing and punctuation.
- Jachnun v1.0: code-switching. The default in the Hebrish apps.
- Sabich v1.0: code-switching plus tech vocabulary, brand names, numbers and the other task recordings.
A real result: the speaker recorded this in this demo (2026-10-15):
Transcript Stock Parakeet v2 I want to go to the Micollette today to buy my groceries. Fine-tune I want to go to the makolet today to buy my groceries. Makolet (מכולת) is the Israeli corner grocery shop. Stock v2 has never heard the word, so it guesses an English-looking name; the fine-tune writes the word itself.
Record or upload audio and all three models transcribe it. Or try one of the held-out test clips below. Both fine-tunes were trained on one speaker's recordings, so try them on your own voice and tell Daniel how they did at hebrish.app or info@hebrish.app.
| Audio | Reference transcript (example clips only) |
|---|
Results (held-out test split, scored on one dataset revision, 2026-10-20)
| Test set | Stock v2 | Jachnun v1.0 | Sabich v1.0 |
|---|---|---|---|
| Code-switching WER | 11.82% | 5.42% | 5.30% |
| Held-out Hebrew terms transcribed correctly | 23.8% | 46.9% | 48.1% |
| Speaker's English WER | 2.81% | 1.56% | 1.62% |
| Tech dictation WER | 2.51% | 1.74% | 1.54% |
| LibriSpeech test-clean WER | 1.23% | 1.20% | 1.30% |
| Long-form WER (~3 min files) | 10.34% | 5.84% | 5.77% |
WER is normalised (lowercase, no punctuation). Differences under about one point are within noise at these test sizes. Example clips here are from the test split and were never trained on. Public benchmark: leaderboard.
Links: hebrish.app · all models · Jachnun · Sabich · dataset
License and attribution: the base model nvidia/parakeet-tdt-0.6b-v2 is by NVIDIA, released
under CC-BY-4.0. The fine-tunes and the example
audio (Daniel Rosehill) are also CC-BY-4.0, used under the Hebrish terms.
Audio is capped at 3 minutes in this demo.