Models · OpenAI · Whisper large-v3-turbo
Whisper Hebrish (large-v3-turbo)
The first Hebrish model, on Whisper large-v3-turbo. Full fine-tune of Whisper large-v3-turbo (809M). Where Hebrish started, in November 2025; the larger sibling of the Small ACFT model.
At a glance
| Base model | openai/whisper-large-v3-turbo (OpenAI) |
| Current version | Version 2, released 2026-10-17 |
| Licence | MIT · use governed by the terms of use |
| Used by | whisper.cpp and Transformers, by hand |
| DOIs | 10.57967/hf/7019 |
Formats and quants
| Repository | Format and files | Version | For | DOI |
|---|---|---|---|---|
danielrosehill/Whisper-Hebrish main | Transformers, whisper.cpp GGML safetensors 3.24 GB; GGML q8_0 874 MB, q5_0 574 MB, f16 1.62 GB · access request | Version 2 | whisper.cpp on desktop, Transformers | 10.57967/hf/7019 |
Hugging Face issues a DOI per repository, so each format can be cited on its own. Repositories, sizes and DOIs checked on the Hub on 2026-10-20.
Versions
| Version | Released | Run | Where | Notes |
|---|---|---|---|---|
| Version 2 current | 2026-10-17 | — | Whisper-Hebrish | Read-Aloud-Voice-Sentences, 1,929 training rows / 215 min |
| Version 1 | 2025-11-17 | — | Whisper-Hebrish @ v1 | 516 desktop clips (about 0.4 h) |
About it
A full fine-tune of openai/whisper-large-v3-turbo (809M parameters). Version 1 (2025-11-17) is where Hebrish started: 516 desktop clips, recorded that day. Version 2 (2026-10-17) was retrained on Read-Aloud-Voice-Sentences, the same parent dataset the Parakeet models use, with canonical spellings and 29% monolingual English to keep its English intact. The output is English with Hebrew words romanized; it does not transcribe Hebrew.
Results from the model card, on a term-disjoint test set (the Hebrew terms in it never appear in training), phone takes only, greedy decoding:
| Metric | Stock turbo | Version 1 | Version 2 |
|---|---|---|---|
| Code-switching WER (normalized), 125 utterances | 9.8% | 7.1% | 4.9% |
| Recall on unseen Hebrew terms (109 occurrences, 70 terms) | 36.7% | 48.6% | 53.2% |
| Own English WER (normalized), 33 utterances | 4.0% | 4.1% | 2.9% |
| LibriSpeech-PC test-clean sample WER (normalized), 200 utterances | 1.8% | 2.2% | 2.2% |
This test set is not the one Parakeet's benchmark uses, so these numbers do not compare with Jachnun's. Trained on one speaker's voice; the test set is small.
Suggested implementations
whisper.cpp (desktop): q8_0 is the recommended build. Always pass -l en; with auto-detection Whisper may switch to Hebrew script.
ffmpeg -i input.m4a -ar 16000 -ac 1 -c:a pcm_s16le clip.wav
whisper-cli -m ggml-whisper-hebrish-v2-large-v3-turbo-q8_0.bin -l en -f clip.wavTransformers:
from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="danielrosehill/Whisper-Hebrish")
print(asr("clip.wav", generate_kwargs={"language": "en", "task": "transcribe"})["text"])For a phone, the Small ACFT model is the one to use: a quarter of the size, and built for FUTO Keyboard.
Cite
Cite Whisper Hebrish with its DOI, 10.57967/hf/7019 (danielrosehill/Whisper-Hebrish). To cite one format, use that repository's DOI from the table above.
@misc{rosehill_whisper_turbo_2026,
author = {Daniel Rosehill},
title = {Whisper-Hebrish},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/danielrosehill/Whisper-Hebrish},
doi = {10.57967/hf/7019}
}Request access
Request access on Hugging Face first. The Whisper Hebrish repositories are gated: open the repository, sign in, and fill in the short request form there. Requests are reviewed by hand. For non-commercial use, the attestation form covers what we ask.
No answer after a few days? Request access here instead, or write to access-requests@hebrish.app. Commercial use, including telephony, needs a commercial licence.