Benchmarks
Every model on the public test sets and the held-out split, each chart with a table and its method.
Technical
Technical
The numbers, the models, the data and the docs. It starts with the latest results: how Sabich has moved across its last three versions on the public benchmark.
Sabich v2.1 to v2.3 on the public benchmark
Sabich is the model the apps recommend. These are its last three versions on test set v2: 129 code-switching clips (14.8 min) across the 13 domains of the lexicon, with 221 Hebrew term occurrences, none of them clips any Hebrish model was trained on. Pick a measure to see how it changed from version to version.
Show as a table
| Model | Code-switching WER | Term recall | Unseen-term recall | Seen-term recall | English control WER | Tech WER |
|---|---|---|---|---|---|---|
| Sabich v2.3 current | 4.96% | 67.0% | 59.9% | 78.6% | 2.42% | 3.02% |
| Sabich v2.2 | 4.91% | 70.1% | 62.0% | 83.3% | 2.12% | 2.76% |
| Sabich v2.1 | 5.21% | 65.6% | 56.2% | 81.0% | 2.42% | 3.02% |
| Parakeet TDT 0.6B v2 (stock, for reference) | 13.72% | 19.9% | 16.1% | 26.2% | 2.73% | 7.54% |
Test set v2, clean subset, normalised WER, greedy decoding; Sabich v2.1 scored 2026-10-25, v2.2 2026-10-27, v2.3 2026-10-30, each on one NVIDIA L40S. From v2.1 to v2.2 every measure improved: code-switching WER 5.21% to 4.91%, unseen terms 56.2% to 62.0%. Sabich v2.3 scores slightly below v2.2 on every measure here (code-switching WER 4.96% against 4.91%), and ahead of it on the held-out split of its own dataset (4.74% against 5.12%); the two test sets disagree, and both differences are small. Against stock Parakeet (13.72%), all three cut code-switching errors by more than half. All models and both test sets: benchmarks.
Where to go next
Models
Sabich, Jachnun, Shakshuka and Meorav: versions, formats, base models, and which one to use.
Datasets
The recordings behind the models, the lexicon of Hebrew terms, and which revision each model was trained on.
Docs
Model card, integration, the local server, the stack, pipelines, the white paper and entry points for AI agents.