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
Log in or Sign Up to review the conditions and access this model content.
Sabich v1.0: ONNX fp32 and int8
ONNX export of Sabich v1.0, the kitchen-sink Parakeet TDT 0.6B v2 fine-tune: English-Hebrew code-switching plus tech vocabulary, brand names, numbers in digits and Irish-English phrasing, in one speaker's voice. Read the main card for results and limitations.
Website: hebrish.app. Use is governed by the Hebrish terms of use: non-commercial use is free, commercial use needs a licence from access-requests@hebrish.app.
The layout is the one used by onnx-asr and
istupakov/parakeet-tdt-0.6b-v2-onnx, exported the
same way (NeMo model.export) from training run r7-all, the same files layout as
Hebrish/Jachnun-ONNX:
| File | Precision |
|---|---|
encoder-model.onnx + encoder-model.onnx.data (2.4 GB) |
fp32 |
decoder_joint-model.onnx |
fp32 |
encoder-model.int8.onnx (652 MB) |
int8, dynamic (QUInt8) |
decoder_joint-model.int8.onnx (9 MB) |
int8, dynamic (QInt8) |
vocab.txt, config.json |
nemo-conformer-tdt, 128 mel features |
Checked 2026-10-21 with onnx-asr 0.12.0 on five held-out phone test clips against the PyTorch (NeMo) output: fp32 identical on 5 of 5, int8 on 3 of 5 (it capitalised "Hi gal" to "Hi Gal" and dropped one comma).
onnx-asr
# pip install onnx-asr[cpu] huggingface_hub
from huggingface_hub import snapshot_download
import onnx_asr
path = snapshot_download("Hebrish/Sabich-ONNX",
allow_patterns=["*.int8.onnx", "vocab.txt", "config.json"])
model = onnx_asr.load_model("nemo-conformer-tdt", path, quantization="int8")
print(model.recognize("audio.wav")) # 16 kHz mono
Drop allow_patterns and quantization for fp32. Voxtype loads this
layout too; give it a directory holding one precision only (instructions on the
Jachnun-ONNX card).
- Downloads last month
- -