# Daniel Rosehill: AI work

Hub: https://ai.danielrosehill.com/ · JSON: https://agents.danielrosehill.com/ai.json

## Focus areas

### English–Hebrew code-switching

Speech and text models that cope with "Hebrish" — the English-with-Hebrew-words that English speakers in Israel actually use. Off-the-shelf ASR mangles words like teudat zehut or mashkanta; I build the datasets and fine-tunes that fix that.

### Speech-to-text & on-device inference on Android

Whisper fine-tunes, ASR evaluations and small models that run locally on a phone. Most of my writing starts as voice, so I care about accuracy on real voice notes, background noise and fast on-device inference.

### Accessibility

Using AI to make technology easier to use: voice-first input instead of typing, TTS that pronounces Hebrew words correctly, and tools that reduce friction for people who work in more than one language.

### Benchmarks & AI safety

Evals for how models behave, not just how they score: does a model lecture the user, repeat propaganda when the premise is loaded, endorse conspiracy theories or assume everyone is American? Plus the speech benchmarks behind my ASR work. Open test prompts on Hugging Face.

### AI for education

AI that teaches rather than just does: tutor-style agents, learning loops for code generation, and My Weird Prompts, a podcast that explains a new topic every day.

## Model cards

### [Hebrish (Jachnun v1.0): Parakeet TDT 0.6B v2 for English–Hebrew code-switching](https://ai.danielrosehill.com/models/parakeet-english-hebrew-codeswitching)

Speech-to-text that gets the Hebrew words right when an English speaker in Israel drops them into English: "pick up some bamba from the makolet".

- Status: v1.0 · Jachnun (released 2026-10-19)
- Hugging Face: https://huggingface.co/Hebrish/Jachnun
- Base: [nvidia/parakeet-tdt-0.6b-v2](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2)
- Licence: CC-BY-4.0
- Code-switching, shared test set, 422 clips (normalised WER): 11.82% → 5.42%
- Hebrew terms held out of training: share correct (324 occurrences, 134 terms): 23.8% → 46.9%
- Speaker's own plain English, 84 clips (normalised WER): 2.81% → 1.56%
- LibriSpeech test-clean, 200 utterances (normalised WER): 1.23% → 1.20%
- Long form, 17 files (normalised WER): 10.34% → 5.84%

### [Hebrish (Sabich v1.0): Parakeet TDT 0.6B v2 for code-switching plus tech dictation](https://ai.danielrosehill.com/models/parakeet-tdt-v2-english-hebrew-all-subsets)

The recommended Hebrish model: everything Jachnun was trained on, plus tech vocabulary, brand names, acronyms, numbers and currency, Irish-English phrasing and the rest of my task recordings.

- Status: v1.0 · Sabich (released 2026-10-20)
- Hugging Face: https://huggingface.co/Hebrish/Sabich (gated)
- Base: [nvidia/parakeet-tdt-0.6b-v2](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2)
- Licence: CC-BY-4.0
- Code-switching, held-out split (normalised WER): 11.82% → 5.30%
- Hebrew terms held out of training: share correct: 23.8% → 48.1%
- Tech dictation (normalised WER): 2.51% → 1.54%
- LibriSpeech (normalised WER): 1.23% → 1.30%

### [Boxty v1.0: Parakeet TDT 0.6B v2 for English and tech dictation](https://ai.danielrosehill.com/models/parakeet-tdt-v2-tech-and-accent)

English dictation without Hebrew, in my voice: model and product names, CLI and cloud terms, brands, acronyms, and numbers and currency in digits.

- Status: v1.0 · Boxty (released 2026-10-20)
- Hugging Face: https://huggingface.co/Hebrish/Parakeet-TDT-0.6B-v2-Tech-And-Accent (gated)
- Base: [nvidia/parakeet-tdt-0.6b-v2](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2)
- Licence: CC-BY-4.0
- Hebrew-free English (normalised WER): 1.79% → 1.04%
- Tech dictation (normalised WER): 2.51% → 1.35%
- LibriSpeech (normalised WER): 1.23% → 1.18%
- Code-switched speech (normalised WER): 11.5% → stock level, on purpose

### [Hebrew-Latin Token Classifier: finds Hebrew written in Latin letters](https://ai.danielrosehill.com/models/hebrew-latin-token-classifier)

Finds Hebrew words written in Latin letters inside English text, so text-to-speech can hand them to a Hebrew voice instead of reading makolet as English. There was no model for this before.

- Status: v1 · released (released 2026-10-15)
- Hugging Face: https://huggingface.co/Hebrish/Hebrew-Latin-Token-Classifier
- Base: [FacebookAI/xlm-roberta-base](https://huggingface.co/FacebookAI/xlm-roberta-base)
- Licence: MIT
- All test sentences (468): 0.928 / 0.953 → 0.941
- Generated (430): 0.928 / 0.948 → 0.938
- Dictated, human-chosen (38): 0.929 / 1.000 → 0.963
- Validation, used for model selection (467): 0.950 / 0.948 → 0.949

### [Gemma 3 270M ASR Paragraph Breaks: paragraphs for dictated text](https://ai.danielrosehill.com/models/gemma-3-270m-asr-paragraph-breaks)

Adds paragraph breaks to flat speech-to-text output and changes nothing else. Small enough to run on a phone after the speech model.

- Status: r2 · first release (released 2026-10-20)
- Hugging Face: https://huggingface.co/danielrosehill/Gemma-3-270M-ASR-Paragraph-Breaks
- Base: [google/gemma-3-270m-it](https://huggingface.co/google/gemma-3-270m-it)
- Licence: Gemma Terms of Use
- Break precision: 0.00 → 0.84
- Break recall: 0.00 → 0.72
- F1: 0.00 → 0.78
- Fallbacks (fewer is better): 26 of 26 → 2 of 26

### [Gemma 3 270M Email Key-Line Bolding: bold the key lines of a dictated email](https://ai.danielrosehill.com/models/gemma-3-270m-email-key-line-bolding)

Puts the key parts of a dictated email in Markdown bold and changes nothing else. Runs after the paragraph-breaks model.

- Status: r2 · preliminary (released 2026-10-21)
- Hugging Face: https://huggingface.co/danielrosehill/Gemma-3-270M-Email-Key-Line-Bolding
- Base: [google/gemma-3-270m-it](https://huggingface.co/google/gemma-3-270m-it)
- Licence: Gemma Terms of Use
- Word precision: - → 0.73
- Word recall: - → 0.54
- Word F1: - → 0.62
- Fallbacks (fewer is better): 6 of 6 → 0 of 6

### [Proximate Pseudonymizer 0.6B: swap identifying details for close, consistent substitutes](https://ai.danielrosehill.com/models/proximate-pseudonymizer-0.6b)

Rewrites a text with its names, places, organisations, dates and contacts swapped for close, plausible substitutes, the same one every time. Proof of concept: it still leaves 9.3% of identifying details unchanged.

- Status: pp-r2 · proof of concept (released 2026-10-21)
- Hugging Face: https://huggingface.co/danielrosehill/Proximate-Pseudonymizer-0.6B
- Base: [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B)
- Licence: Apache-2.0
- Leak: identifying details left unchanged (lower is better): 85.6% → 9.3%
- Consistency: texts where each changed detail got one substitute: 0.997 → 0.990
- Rest of the text preserved (non-identifying words kept): 0.791 → 0.998
- Texts with no details returned unchanged: 0.308 → 0.962

### [Hebrish (Sabich-Qwen3-ASR-1.7B v1.0): Sabich on Qwen3-ASR 1.7B for desktop GPUs](https://ai.danielrosehill.com/models/sabich-qwen3-asr-1.7b)

The most accurate Sabich: the same code-switching and tech data, trained on Qwen3-ASR 1.7B, an ASR model with an LLM decoder, for computers with a GPU.

- Status: v1.0 · Sabich-Qwen3-ASR-1.7B (released 2026-10-21)
- Hugging Face: https://huggingface.co/Hebrish/Sabich-Qwen3-ASR-1.7B (gated)
- Base: [Qwen/Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B)
- Licence: Apache-2.0 (base model); use under the Hebrish terms of use
- English–Hebrew code-switching (normalised WER): 10.0% → 5.4%
- Hebrew terms never seen in training: recall: 32% → 50%
- Hebrew terms seen in training: recall: 40% → 73%
- Speaker's own English (normalised WER): 4.6% → 2.5%
- LibriSpeech test-clean, 200 utterances (normalised WER): 1.2% → 1.35%
- Long form, about 3-minute joined clips (normalised WER): 12.0% → 8.7%

### [Hebrish (Shakshuka v1.0): Nemotron streaming fine-tune for code-switching plus tech dictation](https://ai.danielrosehill.com/models/shakshuka-nemotron-streaming)

The streaming line of Hebrish: text appears while you speak, trained on the same everything-in-it recordings as Sabich.

- Status: v1.0 · Shakshuka (released 2026-10-21)
- Hugging Face: https://huggingface.co/Hebrish/Shakshuka (gated)
- Base: [nvidia/nemotron-speech-streaming-en-0.6b](https://huggingface.co/nvidia/nemotron-speech-streaming-en-0.6b)
- Licence: NVIDIA Open Model License Agreement; use under the Hebrish terms of use
- Public benchmark, 84 clips: Hebrew term recall (offline): 32.6% → 62.2%
- Public benchmark: unseen-term recall (offline): 32.5% → 54.7%
- Public benchmark: WER (offline): 9.5% → 4.7%
- Public benchmark: English control WER (offline): 2.7% → 2.5%
- Code-switching, held-out split, 387 clips (normalised WER, offline): 11.3% → 5.8%
- Code-switching, held-out split (normalised WER, streaming at 560 ms): n/a → 5.65%

## Dataset cards

### [English–Hebrew Code-Switching Sentences](https://ai.danielrosehill.com/datasets/english-hebrew-codeswitching-sentences)

English sentences with everyday Hebrew words and Israeli names written in Latin letters, read aloud on a phone by one English speaker living in Israel.

- Hugging Face: https://huggingface.co/datasets/danielrosehill/English-Hebrew-Codeswitching-Sentences (gated)
- Licence: CC-BY-4.0
- Size (2026-10-15): train 1,248 clips, 120.1 min; test 121 clips, 13.9 min

### [Read-Aloud Voice Sentences](https://ai.danielrosehill.com/datasets/read-aloud-voice-sentences)

Every take recorded for the Hebrish speech models: read-aloud English sentences with Hebrew words, tech vocabulary, names, emails and messages, recorded on a phone and published in batches straight from the capture app.

- Hugging Face: https://huggingface.co/datasets/danielrosehill/Read-Aloud-Voice-Sentences (gated)
- Licence: CC-BY-4.0
- Size (2026-10-21): train 5,807 clips, 642 min; test 536 clips, 62.7 min

### [English–Hebrew Code-Switching ASR Benchmark](https://ai.danielrosehill.com/datasets/english-hebrew-codeswitching-asr-benchmark)

A frozen, versioned test set for scoring how well speech recognition writes the Hebrew words in English speech, "pick up some bamba from the makolet", with stock models and fine-tunes scored the same way on a public leaderboard.

- Hugging Face: https://huggingface.co/datasets/Hebrish/English-Hebrew-Codeswitching-ASR-Benchmark (gated)
- Licence: CC-BY-4.0

### [Hebrew-in-Latin Code-Switching Corpus](https://ai.danielrosehill.com/datasets/hebrew-latin-code-switching)

English sentences with Hebrew words written in Latin letters, every Hebrew span marked token by token, for finding those words in text before text-to-speech mispronounces them. No corpus like it had been published.

- Hugging Face: https://huggingface.co/datasets/Hebrish/hebrew-latin-code-switching
- Licence: MIT

## Hebrish series

As listed on hebrish.app (checked 2026-10-21).

### In the Hebrish apps

On-device dictation on Android, Ubuntu and Windows.

- [Sabich v1.0](https://ai.danielrosehill.com/models/parakeet-tdt-v2-english-hebrew-all-subsets) (Recommended): NVIDIA Parakeet TDT 0.6B v2. 73.8% term recall, 3.3% WER
- [Jachnun v1.0](https://ai.danielrosehill.com/models/parakeet-english-hebrew-codeswitching): NVIDIA Parakeet TDT 0.6B v2. 73.3% term recall, 3.3% WER

### Sabich on other bases

The same training data on bigger and smaller models, for GPUs and wearables. Not in the apps yet.

- [Sabich-Qwen3-ASR-1.7B v1.0](https://ai.danielrosehill.com/models/sabich-qwen3-asr-1.7b) (Desktop GPU): Qwen Qwen3-ASR 1.7B. 5.4% code-switching WER (stock Qwen3-ASR 10.0%)
- [Sabich-Canary-Qwen-2.5B v1.0](https://hebrish.app/models/sabich-canary-qwen) (Desktop GPU): NVIDIA Canary-Qwen 2.5B. 6.9% code-switching WER (Sabich-Qwen3-ASR-1.7B 5.4%), 2.4% on the speaker's own English
- [Sabich-Parakeet-110M v1.0](https://hebrish.app/models/sabich-110m) (Small devices): NVIDIA Parakeet TDT-CTC 110M. Not benchmarked yet
- [Sabich-asr4all -S and -M v1.0](https://hebrish.app/models/sabich-asr4all) (Wearables): FUTO asr4all S and M. 9.1% (-M) and 10.1% (-S) code-switching WER (stock asr4all-m 19.5%)

### Streaming and Whisper

Live text while you speak, and the Whisper fine-tunes where Hebrish started.

- [Shakshuka v1.0](https://ai.danielrosehill.com/models/shakshuka-nemotron-streaming) (Streaming): NVIDIA Nemotron Speech Streaming EN 0.6B. v1.0 not benchmarked yet; PoC-2: 62.2% term recall, 4.7% WER (offline)
- [Whisper Hebrish Small ACFT small-en-en30-acft](https://hebrish.app/models/whisper-small): OpenAI Whisper small.en. 5.8% code-switching WER on its own test set (FUTO runtime)
- [Whisper Hebrish Version 2](https://hebrish.app/models/whisper-turbo): OpenAI Whisper large-v3-turbo. Version 1: 54.1% term recall, 6.5% WER
- [Meorav PoC-1](https://hebrish.app/models/meorav) (POC): Qwen Qwen3-TTS 12Hz 1.7B Base. No benchmark yet; listening verdict pending

### Text

A small model that finds Hebrew written in Latin letters inside English text.

- [Hebrew-Latin Token Classifier v1](https://ai.danielrosehill.com/models/hebrew-latin-token-classifier) (Small model): Meta XLM-RoBERTa base. Span F1 0.941 on terms it never saw in training

## Benchmarks and evals

Registry: https://github.com/danielrosehill/My-Benchmarks · page: https://ai.danielrosehill.com/benchmarks

| Benchmark | Category | Stage | Summary |
| --- | --- | --- | --- |
| [English–Hebrew Code-Switching ASR Benchmark](https://github.com/Hebrish-ASR/English-Hebrew-Codeswitching-ASR-Benchmark) | speech | results | Speech-to-text on English speech that contains Hebrew words: a frozen test set, runner, scorer and public leaderboard, comparing stock models with code-switching fine-tunes. |
| [The PC Bench](https://github.com/danielrosehill/The-PC-Bench) | behaviour | prompts | Does the model scold you? Measures lecturing, moralising, terminology correction and unrequested disclaimers on legitimate requests. 24 mirrored pairs (Western-progressive vs conservative/religious triggers) plus neutral controls. |
| [Propaganda Bait Eval](https://github.com/danielrosehill/Propaganda-Bait-Eval) | behaviour | built | Feeds models loaded political premises across six series (CCP, MAGA, pro-Israel, pro-Palestine, Kremlin, US-progressive) and scores how readily each framing is repeated, with per-series selectivity. |
| [Conspiracy Endorsement Eval](https://github.com/danielrosehill/Conspiracy-Endorsement-Eval) | behaviour | results | Whether a model pushes back on, endorses or encourages 45 conspiracy theories, each asked in a curious and a believer framing. First hand-graded run (on-air-01): 7 models. |
| [Israel–Palestine Framing Eval](https://github.com/danielrosehill/Israel-Palestine-Framing-Eval) | behaviour | results | Whether a model treats mirrored pro-Israel and pro-Palestinian framings of the same events differently: loaded premises, framing pairs and open-neutral questions. First graded run (on-air-01): 6 models. |
| [American Centrism Eval](https://github.com/danielrosehill/American-Centrism-Eval) | behaviour | prompts | Whether a model assumes the user is in the US when nothing says so: geography, units, spelling, law and culture defaults, with controls. |
| [Dan-Eval](https://github.com/danielrosehill/Dan-Eval) | behaviour | built | A composite of the model-behaviour checks I run before adopting a model for an app, giving one scorecard per model. Each component lives in its own repo. |
| [Audio Understanding Bitrate Eval](https://github.com/danielrosehill/Audio-Understanding-Bitrate-Eval-0426) | speech | results | How MP3 bitrate affects transcription accuracy across every audio-input LLM on OpenRouter. |
| [Whisper Fine-Tune Accuracy Eval](https://github.com/danielrosehill/Whisper-Fine-Tune-Accuracy-Eval) | speech | results | WER of personal Whisper fine-tunes vs stock Whisper at each model size. Fine-tuning helped the small models, hurt the large ones, and improved code-switching at every size. |
| [Whisper WPM & Background Noise Eval](https://github.com/danielrosehill/Whisper-WPM-Background-Noise-Eval) | speech | results | How speaking pace and background noise affect ASR word error rate. |
| [Long-Form Audio Eval](https://github.com/danielrosehill/Long-Form-Audio-Eval) | speech | results | STT models on a 27-minute podcast against a professional reference transcript: WER, CER and punctuation accuracy. |
| [One-Shot Transcription Microphone Eval](https://github.com/danielrosehill/One-Shot-Transcription-Microphone-Eval) | speech | results | STT accuracy across a range of microphones recording the same material. |
| [Transcription Cleanup Eval](https://github.com/danielrosehill/Transcription-Cleanup-Eval-1225) | speech | results | Cloud audio-multimodal models transcribing and cleaning up a recording in one step. Subjective grading. |
| [Local ASR STT Benchmark](https://github.com/danielrosehill/Local-ASR-STT-Benchmark) | speech | results | Whisper model sizes in Speech Note on Ubuntu with AMD/ROCm. A rough first-pass comparison. |
| [Hebrew Image Generation Eval](https://github.com/danielrosehill/Hebrew-Image-Generation-Eval) | image | results | Twelve text-to-image models scored pass/fail on rendering Hebrew text correctly. |
| [Gemma 4 Benchmarks](https://github.com/danielrosehill/Gemma-4-Benchmarks-0426) | hardware | results | Gemma 4 E4B vs Qwen3-VL 8B via Ollama on a CPU-only i7-12700F with 64 GB RAM. |
| [Workstation Benchmarks](https://github.com/danielrosehill/Workstation-Benchmarks-0226) | hardware | results | Desktop workstation benchmarks (i7-12700F, RX 7700 XT, 64 GB DDR5, Ubuntu 25.10) with the full system inventory. |
| State Censorship in Asian Models | behaviour | planned | State-aligned censorship and framing in Chinese and other Asian models. |
| [US Partisan Framing Eval](https://github.com/danielrosehill/US-Partisan-Framing-Eval) | behaviour | prompts | Asked a neutral question about a divisive US issue, does the model answer with the popular Democratic or Republican framing, and on foreign policy, hawkish or dovish? Blinded anchor-based grading, 14 topics plus factual controls. |
| Last-Hour News Recall | behaviour | planned | Recall of the past hour's news by models that claim real-time retrieval. |
| Last-24-Hours News Recall | behaviour | planned | Recall of the past day's news by models that claim real-time retrieval. |

## My Weird Prompts

A daily AI-generated podcast. I send in a question as a voice prompt, and an AI pipeline researches it, writes the script and voices the hosts, Corn the sloth and Herman the donkey. Topics run from AI and technology to geopolitics, health and history.

- 5,862 episodes published as of 14 October 2026
- Podtrac (IAB-certified) as of 13 October 2026: 89,978 all-time downloads; 14,847 downloads and 2,274 unique listeners in the last 30 days (1,131 in the US). Podtrac does not count Spotify plays, so these figures understate the total audience.
- https://www.myweirdprompts.com

## Hugging Face collections

- [Code Switching (English and Hebrew)](https://huggingface.co/collections/danielrosehill/code-switching-english-and-hebrew-all-resources-6acfdcb1e11911797af4c3b3)
- [Hebrew ASR](https://huggingface.co/collections/danielrosehill/hebrew-asr-and-audio-to-text-69cbb72271b4fe75732a465a)
- [Whisper Hebrish](https://huggingface.co/collections/danielrosehill/whisper-hebrish)
- [Android STT favourites](https://huggingface.co/collections/danielrosehill/android-stt-favorites)
- [Small ASR models](https://huggingface.co/collections/danielrosehill/small-asr-models)
- [STT evaluations](https://huggingface.co/collections/danielrosehill/stt-evaluations)
- [STT cleanup (text post-processing)](https://huggingface.co/collections/danielrosehill/stt-cleanup-text-postprocessing)
- [My public audio datasets](https://huggingface.co/collections/danielrosehill/my-public-audio-datasets)
- [Hebrew TTS](https://huggingface.co/collections/danielrosehill/hebrew-tts)