Support
Hebrish is free, open, and runs on your own device. It is one person's project, trained on one person's voice. Support would let it learn more voices and stay free.
There is no way to donate yet. When there is, it will be on this page. Until then, the ways to help below make more difference than money would.
Dictation is an accessibility tool. For people with RSI, motor impairments or low vision, it is how they write, not a convenience. It should work for the words they actually say. For many English speakers that includes Hebrew words: the arnona, the kupat holim, Motzei Shabbat. Stock dictation models often get those wrong, and fixing them by hand is the very thing dictation was meant to spare.
Hebrish shows that a small model can learn these words. Making that work for everyone, not just one voice, is the next step.
Every recording so far is one man's voice: an Irish-accented English speaker in Jerusalem. The model needs women's voices, other accents, and people who grew up with Hebrew as well as people who learned it later. Collecting and checking those recordings takes time.
Each training run takes well under an hour on one NVIDIA L40S GPU and costs under a dollar today. More voices and more versions mean more runs, and every version should be scored on the benchmark. Versions 4 and 5 are not benchmarked yet.
The apps are free. Transcription runs on your device, and your audio is not uploaded. The model and dataset are CC-BY-4.0. Keeping it that way means keeping the Android, Ubuntu and Windows apps working without charging for them.
Stars and shares help other people find it. If Hebrish is useful to you, star the repositories, like the model on Hugging Face, and tell someone who dictates.
New model versions, app releases, and calls for voice recordings. Confirm by email; unsubscribe from any message.
The next model versions and the recording targets for each: see the roadmap. Who it is for: use cases.