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.

Sabich-Qwen3-ASR-1.7B v1.0

Sabich trained on a larger base model. Sabich is the kitchen-sink line of the Hebrish project: one speaker's English dictation with everything in it. That covers Hebrew words written in canonical romanization (teudat zehut, arnona, makolet), tech vocabulary, brand names, acronyms, numbers and currency in digits, and everyday email and message dictation. The main Sabich is Parakeet TDT 0.6B v2, sized for phones. This variant applies the same training data to Qwen/Qwen3-ASR-1.7B, an ASR model with an LLM decoder, for desktop GPUs and the highest accuracy.

Name. Each line has a food code name for what it is trained on. A variant on a different base model keeps the line name and adds the base: Sabich-<Base>-<Size>.

Output

English, punctuated and cased. Hebrew words are written in Latin script, never Hebrew script. Always transcribe with language="English", which is the prefix the model was trained with.

Results

Held-out test split of the speaker's recordings (term-disjoint from training), dataset revision cf84f7e4. WER is normalised (lower-cased, punctuation stripped); term recall is a word-boundary match on the romanized Hebrew term.

	Stock Qwen3-ASR-1.7B	This model
English-Hebrew code-switching, 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, WER	4.6%	2.5%
LibriSpeech test-clean (200 utterances), WER	1.2%	1.35%
Long-form (about 3-minute joined clips), WER	12.0%	8.7%
Numbers written in digits	2/8	8/8

Use

import torch
from qwen_asr import Qwen3ASRModel  # pip install qwen-asr

model = Qwen3ASRModel.from_pretrained("Hebrish/Sabich-Qwen3-ASR-1.7B", dtype=torch.bfloat16, device_map="cuda:0")
print(model.transcribe(audio="note.wav", language="English")[0].text)

About 5 GB of VRAM in bf16. On AMD GPUs it runs under PyTorch ROCm.

Run it without PyTorch

GGUF builds are in Hebrish/Sabich-Qwen3-ASR-1.7B-GGUF:

llama.cpp/ for desktops: llama-mtmd-cli or llama-server with --mmproj, on a GPU through HIP (ROCm), Vulkan or CUDA, or on a CPU. bf16, Q8_0 and Q4_K_M.
transcribe.cpp/ for Android: import the Q4_K_M file into Offline Voice Input (dev.notune.transcribe).

The GGUF card has the exact commands and a check of every file against this model.

Training

Data: danielrosehill/Read-Aloud-Voice-Sentences at revision e918e381, blend all: 592 minutes of the speaker's own takes (383 of them code-switching), plus 10% LibriSpeech-PC replay against forgetting. 658 minutes in total.
Recipe: the official Qwen3-ASR full fine-tuning script (finetuning/qwen3_asr_sft.py, upstream commit 7c6daf7). 3 epochs, lr 2e-5 with linear decay, 5% warmup, effective batch 32, bf16, one H100, 15 minutes. The epoch-3 checkpoint was chosen.
Targets: language English<asr_text> + the punctuated, cased transcript.

Limits

Trained on one speaker (an Irish English speaker living in Israel), reading prompts on a phone. Other voices will gain less.
About half of unseen Hebrew words are still missed or misspelled.
Long-form audio is its weakest area.

Licence: the base model is Apache-2.0; use of this fine-tune is governed by the Hebrish terms of use.