You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Access to this model is by request and is governed by the Hebrish terms of use (https://hebrish.app/terms). Non-commercial use is free; commercial use needs a commercial licence from access-requests@hebrish.app. Please also complete the attestation form: https://hebrish.app/legal/hebrish-terms-of-use.pdf

Sabich v1.0 (GGUF): English-Hebrew code-switching plus tech

GGUF builds of Sabich v1.0, a fine-tune of NVIDIA Parakeet TDT 0.6B v2 trained on everything: English with romanized Hebrew words dropped in ("arnona", "makolet", "teudat zehut") plus tech vocabulary, brand names, acronyms, numbers and currency. It is the kitchen-sink sibling of Jachnun v1.0: same base, retrained from stock on a bigger blend, not trained on top of Jachnun. Read the main model card for the training data and full results.

Website: hebrish.app: the Hebrish apps for Android and desktop (Sabich is in their model picker), the results, and every model and dataset. Android updates through F-Droid: add the Hebrish F-Droid repository (https://fdroid.hebrish.app/repo; steps).

Access and terms of use. Use of this model and every converted or quantised build of it (GGUF, ONNX, sherpa-onnx) is governed by the Hebrish terms of use. Access is by request: use the button on this page, or write to access-requests@hebrish.app. Non-commercial use (personal, research, education, open-source projects) is free; please complete the attestation form. Commercial use, including telephony and call centres, needs a commercial licence: access-requests@hebrish.app. Versions released under CC-BY-4.0 (up to v1.0) stay under that licence; the terms do not restrict rights it grants.

Converted with transcribe.cpp from training run r7-all (Sabich v1.0, 2026-10-20). Same files as gguf/ in the main repo, copied here so the GGUF builds of both Hebrish models sit side by side.

File	Size	Use
`…-Q8_0.gguf`	730 MB	Recommended. Phones and desktop CPUs
`…-Q4_K_M.gguf`	475 MB	Phones short on memory. Slower than Q8_0 on x86
`…-F16.gguf`	1.2 GB	GPU inference (transcribe.cpp built with Vulkan, HIP/ROCm or CUDA)

Check against full precision (5 held-out phone clips, 2026-10-20, against the PyTorch model): F16 5/5 identical, Q8_0 4/5, Q4_K_M 2/5. Use Q8_0 unless memory forces Q4_K_M.

Results

On the public benchmark (scored with the main model), Sabich v1.0 writes 73.8% of Hebrew terms exactly (Jachnun v1.0 73.3%, stock Parakeet v2 32.6%) at 3.3% WER. On the held-out split it is level with Jachnun on code-switching (5.30% vs 5.42% WER) and better on tech dictation (1.54% vs 1.74%). Full results: hebrish.app/benchmarks.

Apps

Android: Offline Voice Input: import the Q8_0 file through its file picker. Or use the Hebrish app, which downloads Sabich itself.
Linux / macOS / Windows: Handy: put the Q8_0 file in Handy's models directory (on Linux ~/.local/share/com.pais.handy/models/) and restart Handy.
Command line: transcribe-cli -m Parakeet-TDT-0.6B-v2-English-Hebrew-Codeswitching-All-Subsets-Q8_0.gguf audio.wav (16 kHz mono WAV).

Other formats: sherpa-onnx int8 (what the Hebrish apps use) · NeMo original.

Licence: CC-BY-4.0, as for the base model (NVIDIA Parakeet TDT 0.6B v2).