Sabich-asr4all-S and -M wearables

At a glance

Formats and quants

Hugging Face issues a DOI per repository, so each format can be cited on its own. Repositories, sizes and DOIs checked on the Hub on 2026-10-21.

Versions

About Sabich-asr4all

Sabich-asr4all-S v1.0 (28.2M parameters) and Sabich-asr4all-M v1.0 (55.6M), 2026-10-21, put Sabich's training data on futo-org/asr4all-s and futo-org/asr4all-m, streaming CTC models with a punctuation and casing head (PCEC). They run on the CPU. -M is about one point better than -S on most tests, at twice the size. Neither is in the Hebrish apps yet.

On the held-out test split of the speaker's recordings at dataset revision cf84f7e4 (term-disjoint from training; streaming tier c256r32, PCEC on; normalised WER unless marked raw):

Runs a1-all (-M, best epoch 15) and a1s-all (-S, best epoch 18): Sabich's data (blend all) at revision cf84f7e4, 5,748 clips with 10% LibriSpeech-PC replay. The CTC trunk was fully fine-tuned, 20 epochs on one L40S, each step on one of the six shipped streaming tiers so that every tier stays trained; the PCEC head was not retrained. These numbers are on a different test-set revision from Sabich's, so the models have not been compared head to head.

Run it

Transformers (4.57, trust_remote_code) on the CPU, with 16 kHz mono audio. The model card explains the streaming tiers; op=(256, 32) is the tier scored above:

import numpy as np, soundfile as sf, torch
from transformers import AutoModel

model = AutoModel.from_pretrained("Hebrish/Sabich-asr4all-M", trust_remote_code=True).eval()
audio, sr = sf.read("note.wav", dtype="float32")
audio = np.concatenate([audio, np.zeros(16000, dtype=np.float32)])  # 1 s of silence keeps the last word
print(model.transcribe(torch.from_numpy(audio), op=(256, 32)))

• ONNX: each repository's onnx/ folder has one graph per streaming tier, the PCEC head and the VAD, sharing one weights file (108 MB for -S, 214 MB for -M). Run it with the included onnx/asr4all_onnx.py and ONNX Runtime. There is no ExecuTorch build.
• How to use them, including the lower-latency tiers: the model cards, Hebrish/Sabich-asr4all-S and Hebrish/Sabich-asr4all-M.
• Output is English, cased and punctuated by the stock PCEC head, with Hebrew words in Latin letters, never Hebrew script.

Limitations

• Numbers come out as words ("four point one", not "4.1"). Their vocabulary has no digits, so the training transcripts spell numbers out.
• Punctuation and casing come from the stock PCEC head, which was not retrained; that is the gap between raw and normalised WER.
• LibriSpeech got worse than the stock model (2.2% stock -M; 4.0% -M, 6.2% -S): the training mix had only 10% general English.
• About two thirds of the Hebrew words they never saw in training are still missed or misspelled. For more accuracy on a small device, Sabich-Parakeet-110M is the next size up.

Licence

Both models are derived from Asr4All (futo-org/asr4all-s and futo-org/asr4all-m) and are subject to the FUTO Model Weights License 1.0, included in each repository as LICENSE. The weights have been modified: the CTC trunk was fine-tuned; the PCEC head is unmodified. Use is also governed by the Hebrish terms of use. These models are not affiliated with or endorsed by the licensor.

Cite

Cite Sabich-asr4all with its DOI, 10.57967/hf/11177 (Hebrish/Sabich-asr4all-S). To cite one format, use that repository's DOI from the table above.

@misc{rosehill_sabich_asr4all_2026,
  author    = {Daniel Rosehill},
  title     = {Sabich-asr4all-S},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/Hebrish/Sabich-asr4all-S},
  doi       = {10.57967/hf/11177}
}

Request access

Which model should I use? · All Sabich variants · All Hebrish models