Open Source

Hebrish: English–Hebrew Code-Switching

English speakers who have moved to Israel tend to talk in a hybrid: English sentences carrying everyday Hebrew words — makolet, teudat zehut, Bituach Leumi. Stock speech recognition models transcribe these phonetically, and text-to-speech reads them as English. I'm building the data and models to close that gap for immigrant speakers, starting with Hebrew and in a way that carries over to other heritage-language communities.

Hebrish is the first result: on-device English dictation that writes the Hebrew words correctly, in romanized form. It is a fine-tune of NVIDIA Parakeet TDT 0.6B v2, trained on my own recordings. It runs in the Hebrish apps: an Android voice keyboard plus tray apps for Ubuntu and Windows. There are two models in the apps: Jachnun v1.0 for code-switching, and Sabich v1.0, which adds tech vocabulary and other dictation data. Both are retrained from stock Parakeet, not stacked on each other. Their results, and the other model lines, are in the ASR fine-tunes catalogue. It is a worked example of fine-tuning a speech model on a small custom dataset and shipping it on-device.

• ASR fine-tunes catalogue Every model line (Jachnun, Sabich, Boxty and more), what each is for and how they differ
• hebrish.app Project website
• Parakeet-TDT-0.6B-v2-English-Hebrew-Codeswitching Model weights for Jachnun v1.0
• English-Hebrew-Codeswitching-Sentences Read-aloud audio with a term-disjoint test split
• English-Hebrew-Codeswitching-Parakeet-Demo Try the model in the browser
• Sabich-sherpa-onnx-int8 Sabich v1.0, the build the apps run
• Hebrish-App The Android, Ubuntu and Windows apps
• PoC-3 write-up How an earlier proof-of-concept checkpoint (published as v3) was trained and evaluated
• hebrew-latin-token-classifier Finds romanized Hebrew inside English text (span F1 0.94)
• hebrew-latin-code-switching Token-level BIO corpus the classifier is trained on
• Whisper-Hebrish Earlier Whisper Large v3 Turbo fine-tune on code-switched speech

Voice-Note Intent Classification

A voice note might be a shopping list, a to-do, a calendar entry or an email draft. In a voice-to-task workflow, something has to decide which before an agent can act on it. I'm developing a small multi-label classifier that reads the transcript and routes it to the right subagent and MCP integration — the classifier decides where a note goes, and never performs the action itself. The label taxonomy is published as a draft ahead of the weights.

• Voice-Note-Intent-Classifier Project plan and draft taxonomy; weights to follow
• Voice-Note-Audio Real-world voice notes with transcripts and annotations
• STT-Voice-Notes-Evals Speech-to-text evaluation on everyday voice notes

IFVI Global Value Factors Database

The International Foundation for Valuing Impacts publishes the Global Value Factor Database: multipliers that convert companies' environmental and social impacts into US dollars, so they can sit alongside financial reporting. I restructured it for machine readability and released it as an open dataset on Hugging Face — the first time it was available there — so it can go straight into analysis and visualization work.

• Global-Value-Factor-Database-Refactor-V2 Current version, August 2025
• ifvi_valuefactors_deriv First release, December 2024 (superseded by V2)

Model Behaviour & AI Safety Evals

Choosing a model for a client deployment means knowing how it behaves, not just how it ranks on public leaderboards. A model that lectures your customers, repeats a government's talking points when a question is loaded, or assumes every user is American can be a liability in production. So I build small, focused evals for the behaviours that matter when shortlisting a model, and publish the test prompts openly.

Each eval uses mirrored or counterweighted prompts, so it can tell a model that leans one way from one that is simply cautious about everything.

• The PC Bench Does the model scold the user? Moralising, lecturing and unrequested disclaimers (prompts)
• Propaganda-Bait-Eval How readily a model repeats loaded political framings, across six series (prompts)
• Conspiracy-Endorsement-Eval Pushback vs endorsement on 45 conspiracy theories (prompts)
• American-Centrism-Eval Does the model assume the user is in the US? (prompts)
• English–Hebrew code-switching ASR leaderboard A replicable speech-to-text benchmark with a frozen test set
• All benchmarks The full index of my benchmarks and evals, with repos and datasets