Daniel Rosehill Blog
Exploring technology, AI, business insights, and personal reflections from a curious mind
Hebrish v1.0: benchmarked
The sixth training run of my English-Hebrew dictation model is v1.0, the first release I will stand by. It writes 73.3% of Hebrew terms correctly on the public benchmark, against 32.6% for stock Parakeet, and English did not get worse.
My Weird Prompts passes 6,000 episodes
My Weird Prompts, my fully automated AI podcast, crossed 6,000 episodes on 15 October 2026: 5,991 in the archive, 2,688 hours of audio, and new translated editions in eight languages.
Voice typing that understands Hebrish
Speech-to-text mangles the Hebrew words English speakers in Israel use every day. I taught an open model to get them right. What code-switching is, real examples, free apps for Android, Linux and Windows, and how you can help.
Parakeet English-Hebrew code-switching model, now in MLX for Apple Silicon
An MLX port of v3 of my English-Hebrew code-switching Parakeet fine-tune, so it runs natively on Apple Silicon Macs via parakeet-mlx or mlx-audio. Same 4.4% WER as the NeMo original on the check set.
Parakeet English-Hebrew Fine-Tune: v3 Released
v3 of my English-Hebrew code-switching Parakeet fine-tune is out. Code-switching WER drops from 10.9% (stock) to 5.3%. Charts compare every version on the same test.
Popular tags
All tags →
English-Hebrew Code-Switching Parakeet: Version 3
Version 3 of an English-Hebrew code-switching fine-tune of NVIDIA Parakeet TDT 0.6B v2: results against stock and earlier versions, what more data did and did not improve, and downloads.
Whisper Hebrish Small ACFT: a code-switching model for FUTO Keyboard
A phone-sized Whisper Hebrish for FUTO Keyboard on Android. It halves the stock English model's error rate on English-with-Hebrew dictation (11.8% to 5.8% WER) and imports as a single 264 MB file.
Whisper Hebrish v2 Released
Version 2 of Whisper Hebrish, my Whisper fine-tune for English with Hebrew words mixed in. Code-switching WER drops from 14.7% (stock turbo) to 6.7%, and there are whisper.cpp builds.
inventory-spec-extractor: a 34 MB model for reading product spec lines
A small token-classification model that pulls voltages, connectors, materials and 22 other attributes out of one line of a product spec. Runs on CPU in about 20 ms per line. F1 0.85 against a regex baseline's 0.58.
Can speech-to-text spell ‘makolet’? I tested seven models on Hebrish
English speakers in Israel talk Hebrish, and dictation software mangles the Hebrew words. I built an open benchmark to measure it: my fine-tune gets two-thirds of them right, off-the-shelf models a third to under half.