Falafel: a Hebrish model for voice commands
The third model in active development, beside Sabich and Shakshuka. Its first release is training now.
Falafel is the new Hebrish model, and the third one in active development, beside Sabich and Shakshuka. Sabich and Shakshuka are for dictation. Falafel is for voice commands: the short things you say to a computer or phone to make something happen. Its first release is training now.
What it is for
Dictation and commands are different jobs. Dictation is sentences that end up as text: a message, a note, an email. A command is a few words that should do something, and it has to be heard right the first time. Falafel is trained for that second job:
- opening apps and controlling the desktop;
- the smart home;
- sending messages and telling AI agents what to do.
Like the rest of Hebrish, it writes English with the Hebrew words in Latin letters, so a command with a Hebrew word in it comes out the way it was said.
Two sizes
- Falafel, on NVIDIA Parakeet TDT 0.6B v2, for desktops.
- Falafel 110M, on NVIDIA's much smaller Parakeet TDT-CTC 110M, for phones and laptop CPUs.
Both ship as sherpa-onnx int8 offline transducers.
Training data
Falafel is trained on Daniel's own recordings of voice commands, the Voice-Command-Sentences dataset: about 5,500 takes, about 4.2 hours, when the first runs started. On top of that, 20% of the training mix is general English, so the model doesn't fall over on something that isn't on the list.
Where it stands
The first training runs started on 2026-11-10. The first release will be Falafel v1.0 and Falafel 110M v1.0. It is not out yet, and the model page will show the version once it is.
Falafel is not benchmarked, and no benchmark is planned for now, so there are no accuracy figures for it. The best test is using it: say what it gets right and wrong through feedback.
In the apps
The Hebrish apps already have a place for it. Android 1.028 and Ubuntu 1.028 (2026-11-10) list Falafel under Models, in the full size and the small one. It shows Coming soon until the first version is out, then downloads like any other model, with no app update needed.
On Hugging Face
Four repositories in the Hebrish organisation will hold the weights: Hebrish/Falafel, Hebrish/Falafel-sherpa-onnx-int8, Hebrish/Falafel-Parakeet-110M and Hebrish/Falafel-Parakeet-110M-sherpa-onnx-int8. Like the other Hebrish models they are gated, with each request reviewed by a person, and licensed CC-BY-NC-4.0. Questions about access go to access@hebrish.app.
