Hebrish docs

At a glance

Current model

v1.0 · Jachnun

Released

2026-10-19

Base model

NVIDIA Parakeet TDT 0.6B v2

Code-switched training audio

339 minutes

Speakers

One, so far

Licence

CC-BY-4.0 (to v1.0)

Model

• Model card · formats, version history, training recipe, limits
  • All models · Parakeet, the Nemotron streaming variant, Whisper Small for FUTO Keyboard
  • Benchmarks · public benchmark and held-out split, every chart with a table
  • Changelog · model versions and app releases
  • Roadmap · live dataset counts and the next checkpoints
• White paper · v1.0: dataset, training, exports, evaluation, limits
  • English (PDF)
  • עברית (PDF)

Data

• Datasets · every dataset, and which revision each model was trained on
  • Code-switching dataset · what was recorded and how, splits, schema
  • Clip archive · 40 recordings to play and search
• Lexicon · every Hebrew word and name covered, by domain
  • lexicon.json · lexicon.csv
  • Send us words · missing or wrongly transcribed

Build with it

• Integrate · the dataset and lexicon in any ASR pipeline
• Stack · what Hebrish is built with
• Other runtimes · GGUF, MLX, ONNX, NeMo
• Terms of use · free for non-commercial use; commercial licence for products

For AI agents

Hebrish publishes machine-readable entry points, so an agent can answer questions about it without scraping pages.

• llms.txt · summary and page index
• Agent skills · discovery index (Agent Skills discovery RFC v0.2.0)
  • hebrish-lexicon-lookup · how Hebrish spells a Hebrew word or name
  • hebrish-model-facts · versions, benchmark numbers and how to quote them
  • hebrish-dataset-usage · training or evaluating on the dataset
• WebMCP tools · in browsers with WebMCP, every page registers read-only tools
  • lookup_hebrew_term · get_model_results · get_downloads · get_project_status · search_sample_clips · open_page
• Raw JSON · CORS open
  • results.json · benchmarks; targets are goals, not results
  • stats.json · dataset counts, refreshed daily
  • lexicon.json
• sitemap.xml · sitemap (readable)

Elsewhere

• Hebrish on Hugging FaceModels, datasets, Spaces
• Hebrish on GitHubCode and tooling
• Model on Hugging Face.nemo, GGUF, ONNX, sherpa-onnx
• Dataset on Hugging FaceRecordings with transcripts
• BenchmarkFrozen test set, models side by side
• Hebrish appSource code
• TrackerWhere the project stands
• SpellingsThe romanization every transcript follows
• Write-upHow PoC-3 was trained