Hebrish
Technical

Models · NVIDIA · Parakeet TDT 0.6B v2 and Parakeet TDT-CTC 110M

Falafel voice commands

The voice-command model: short spoken commands to a computer or phone, in English with the Hebrew words written in Latin letters. Open an app, control the desktop or the smart home, send a message, tell an AI agent what to do.

In active development, with Sabich and Shakshuka. Falafel is one of the three Hebrish models being worked on, beside Sabich (the default for voice typing) and Shakshuka (streaming). The first release, Falafel v1.0 and Falafel 110M v1.0, is training now (since 2026-11-10). The Hebrish apps already list it, and it arrives in them on its own once it is out.

first release training now not benchmarked

What it is for

Sabich and Shakshuka are built for dictation: sentences, messages, notes. Falafel is built for the other thing people say to a device: a few words that should make something happen. Commands are short, they lean on names, and they mix in Hebrew words the same way dictation does. Falafel is trained on exactly that.

  • Computer and phone: opening apps, controlling the desktop.
  • Smart home: controlling devices by voice.
  • Messages and agents: sending a message, giving an instruction to an AI agent.

Two sizes

ModelBase modelForIn the apps now
Falafelnvidia/parakeet-tdt-0.6b-v2DesktopNot released yet
Falafel 110Mnvidia/parakeet-tdt_ctc-110mPhones and laptop CPUsNot released yet

Both ship as sherpa-onnx int8 offline transducers. The last column is read from the apps' model manifests (cdn.hebrish.app/models/falafel/latest.json and cdn.hebrish.app/models/falafel-110m/latest.json), so it shows the version as soon as one is out.

What it 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), plus 20% general English, so it still copes with anything that isn't on the list.

In the Hebrish apps

Android 1.028 and Ubuntu 1.028 (both 2026-11-10) list Falafel under Models, in the full size and the small one (changelog). It shows Coming soon until the first version is out, then downloads like any other model, with no app update.

Evaluation

Falafel is not benchmarked, and no benchmark is planned for now. There are no accuracy figures for it on this site; say what it gets right and wrong through feedback.

At a glance

Base modelnvidia/parakeet-tdt-0.6b-v2 (NVIDIA)
Current versionNot released yet: Falafel v1.0 and Falafel 110M v1.0 are training (since 2026-11-10)
Parameters0.6B / 114M (base models' sizes: Falafel, Falafel 110M)
EvaluationNot benchmarked, and no benchmark is planned for now
Training data: Voice-Command-Sentences
AccessNot published yet
Downloadevery format and quant (Hugging Face)
LicenceCC-BY-NC-4.0, like Sabich · use governed by the terms of use
Used byThe Hebrish apps (Android 1.028 and Ubuntu 1.028 list it under Models as Coming soon; it downloads once the first release is out); your own tools (sherpa-onnx)
DOIsnone yet

Formats

RepositoryFormat and filesVersionForDOI
Hebrish/Falafel coming soon main.nemo (NeMo)
arrives with v1.0 · not published yet
v1.0 (training)NVIDIA NeMonone yet
Hebrish/Falafel-sherpa-onnx-int8 coming soonsherpa-onnx int8
offline transducer; arrives with v1.0 · not published yet
v1.0 (training)Desktop: Linux, Windows; the Hebrish appsnone yet
Hebrish/Falafel-Parakeet-110M coming soon.nemo (NeMo)
Falafel 110M; arrives with v1.0 · not published yet
Falafel 110M v1.0 (training)NVIDIA NeMonone yet
Hebrish/Falafel-Parakeet-110M-sherpa-onnx-int8 coming soonsherpa-onnx int8
Falafel 110M; offline transducer; arrives with v1.0 · not published yet
Falafel 110M v1.0 (training)Phones and laptop CPUs: Android, Linux, Windows; the Hebrish appsnone yet

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 2026-11-10.

Versions

Not published yet.

Cite

Not published yet. Falafel gets a citation and DOI once it is on Hugging Face.

Request access

The four Hugging Face repositories, Hebrish/Falafel, Hebrish/Falafel-sherpa-onnx-int8, Hebrish/Falafel-Parakeet-110M and Hebrish/Falafel-Parakeet-110M-sherpa-onnx-int8, are gated: a person reviews each request. The weights arrive with v1.0. Questions: access@hebrish.app. Using Falafel in the Hebrish apps needs no access request.

Not published yet.

All Hebrish models