The model in the Hebrish apps (the default since 2026-11-02, updated on a rolling basis from the previous weights; older releases stay on Hugging Face). Trained on everything: the code-switched speech plus tech vocabulary, brand names, numbers and other task recordings.
Current
Sabich v2.3 (2026-10-30)
Last updated
Formats
.nemo, sherpa-onnx int8, ONNX fp32 and int8, GGUF, MLX
Parameters
0.6B
Result
v2.3 on public test set v2 (2026-10-30): 4.96% code-switching WER (stock 13.72%), 67.0% term recall, 59.9% unseen-term recall; slightly below v2.2 there (4.91%, 70.1%), ahead of it on the held-out split
The streaming line: text appears while you speak. Shakshuka v2.1 is trained on Sabich's data (code-switching plus tech) on NVIDIA's cache-aware streaming model.
Current
Shakshuka v2.1 (2026-11-05)
Last updated
Formats
.nemo, sherpa-onnx int8 (streaming), GGUF
Parameters
0.6B
Result
Not benchmarked since PoC-2 (v1.0, v2.0 and v2.1 neither); PoC-2: 62.2% term recall, 4.7% WER on the public benchmark (offline)
All of Sabich's training on NVIDIA's smallest Parakeet, about a fifth of the size. For low-end phones, watches and wearables. Not in the Hebrish apps yet.
Sabich's training data on the tiny asr4all streaming models, 28M and 56M parameters, on the CPU. For wearables and very low-power boards. Numbers come out as words. Not in the Hebrish apps yet.
Current
Sabich-asr4all-S and -M v1.0 (2026-10-21)
Last updated
Formats
asr4all (Transformers remote code, fp32) + ONNX
Parameters
28.2M
Result
9.1% (-M) and 10.1% (-S) code-switching WER (stock asr4all-m 19.5%) on the held-out split at revision cf84f7e4
A proof of concept: Sabich's training data on Qwen3-ASR 1.7B, an ASR model with an LLM decoder, trained once to compare a Qwen3-ASR base with the Parakeet base. Not retrained and not in the Hebrish apps; for everyday use, Sabich is the default in the apps.
Current
Sabich-Qwen3-ASR-1.7B v1.0 (2026-10-21)
Last updated
Formats
Transformers (safetensors, bf16), GGUF
Parameters
2.0B
Result
5.4% code-switching WER (stock Qwen3-ASR 10.0%) on the held-out split at revision cf84f7e4; not run on the public benchmark, so not comparable with Sabich's test set v2 results
Sabich's training data on NVIDIA Canary-Qwen 2.5B, a speech encoder feeding Qwen3-1.7B. For desktop GPUs: the strongest of the Sabich variants on plain English, behind Sabich-Qwen3-ASR-1.7B on code-switching. Not in the Hebrish apps yet.
Current
Sabich-Canary-Qwen-2.5B v1.0 (2026-10-21)
Last updated
Formats
NeMo SALM (safetensors, bf16)
Parameters
2.56B
Result
6.9% code-switching WER (Sabich-Qwen3-ASR-1.7B 5.4%), 2.4% on the speaker's own English, on the held-out split at revision cf84f7e4
Not a Hebrish model, but it comes from the same work: Daniel's own dictation model for English with no Hebrew in it. It has its own licence, and the Hebrish terms don't cover it.
Not part of the Hebrish project: Parakeet plus tech vocabulary. It's the model Daniel uses when there is no Hebrew in what he is saying, a personal fine-tune for English dictation with tech vocabulary, brand names and numbers in digits.
Current
Fritter v1.2 (2026-11-05)
Last updated
Formats
.nemo, GGUF, ONNX, sherpa-onnx int8
Parameters
0.6B
Result
v1.2 not benchmarked yet. v1.0 on its own test clips (not comparable with the Hebrish test sets): 4.01% tech WER (stock 7.53%), 4.73% Hebrew-free English WER (stock 7.46%)
