Hebrish
Experimental Hebrish is an early-stage project, trained on a small code-switching dataset in one speaker's voice. First results are promising. Growing the training data is a long-term effort, and the aim is for every version to get better. Roadmap

Stack

What Hebrish is built with

Hebrish is a small, self-funded project. It stands on open models and tools made by others, and rents a GPU only for the minutes it needs. This is everything it uses, and what for.

Base models

Training and evaluation

  • Modal

    Rented GPUs for every training run and every evaluation.

    One NVIDIA L40S per job. A Parakeet training run takes about 40 minutes and costs under a dollar.

  • NVIDIA NeMo

    Fine-tuning and evaluating Parakeet and Nemotron.

    NeMo 2.7.3.

  • Hugging Face Transformers

    Fine-tuning Whisper.

    Then FUTO's ACFT procedure for the FUTO Keyboard build.

  • The benchmark

    A frozen test set and scorer, run on Modal like training.

    Open source: runner and scorer.

Data and hosting

  • Hugging Face Hub

    Every model, dataset, the leaderboard and the demo Space.

    The apps download the model from here.

  • Claude

    Writing most of the sentences that get read aloud.

    Each generated sentence is reviewed by a second model from a different family. Every recording is a person reading it.

  • GitHub

    App source code, the lexicon and the project tracker.

  • Vercel

    Hosts this website.

    Plain HTML and CSS, deployed on every push.

Running on your device

  • sherpa-onnx

    The engine inside the Hebrish apps on Android, Ubuntu and Windows.

    From the k2-fsa project. Int8 model, CPU.

  • ONNX Runtime

    Runs the model in your browser on /demo, via parakeet.js.

    WebAssembly build.

  • transcribe.cpp and whisper.cpp

    GGUF and GGML builds, for apps such as Offline Voice Input, Handy and FUTO Keyboard.

  • MLX

    The Apple Silicon build, with parakeet-mlx.

    Still PoC-3.

Logos are trademarks of their owners, shown to say what Hebrish uses. Hebrish is not affiliated with or endorsed by any of them. Brand glyphs from Simple Icons (CC0).