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

Projects

What we are working on

Hebrish's projects in one place: what is being considered, what is under way, and what has shipped. Each card links to the details. Where a project needs money, the card says how much and what for.

Under consideration 4

  • Hebrish for iPhone

    A proof-of-concept iPhone app running the same model on the phone, installable from a public TestFlight link. The tasks are listed in the post.

    Funding needed: about $199 plus testing costs. $99 for Apple's developer account (yearly), $100 towards the time. Breakdown

  • Live text with the streaming model

    The Nemotron fine-tune can show text while you speak. It is only a .nemo file today; it needs an on-device export and a test at small chunk sizes before an app can use it.

  • MLX build for Macs, brought up to v1.0

    The Apple Silicon (MLX) build is still PoC-3. Every other format is v1.0.

  • v1.1 retrain

    Planned at about 450 minutes of code-switched speech. The roadmap tracks progress.

In progress 3

Done 7

  • Model v1.0 · Jachnun

    Checkpoint 6 (training run r6-en30), 2026-10-19. The first release past proof of concept. Benchmarked: 73.3% term recall.

  • Benchmark of v1.0 and Nemotron

    2026-10-19, about 50 cents of GPU time. v1.0: 73.3% term recall, 3.3% WER on the public benchmark.

  • Streaming fine-tune on Nemotron

    Checkpoint 2 (run r6-nemotron-en30), 2026-10-19. Experimental: same data and recipe, streams in 80 to 1,120 ms chunks.

  • Whisper Small for FUTO Keyboard

    A 264 MB Whisper fine-tune that FUTO Keyboard on Android can import, 2026-10-17. Halves FUTO's stock error rate on code-switched speech, on its own test set.

  • Hebrish 0.2.0 apps

    Android voice keyboard, Ubuntu and Windows tray apps.

  • Try it in the browser

    The model runs in your browser, on your device. Nothing is uploaded.

  • Clip archive

    Every recording in the dataset, playable, each with an ID you can cite.

Funding a project coming soon

Donations are not open yet. When they are, each project that needs money will show its goal here, so you can see exactly what a donation pays for. Until then, see other ways to help.

Updated 2026-10-19. Release notes for finished work are in the changelog; dataset progress is on the roadmap.