Sabich-Rolling / README.md
danielrosehill's picture
Model card: add the video on how the training data is recorded
9b867e0 verified
|
Raw History Blame
5.54 kB
---
license: cc-by-nc-4.0
language:
- en
- he
base_model: nvidia/parakeet-tdt-0.6b-v2
base_model_relation: finetune
pipeline_tag: automatic-speech-recognition
library_name: nemo
tags:
- parakeet
- tdt
- english
- hebrew
- code-switching
- dictation
extra_gated_prompt: 'Access is approved automatically once you accept the terms. Use
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@hebrish.app.'
extra_gated_fields:
Name: text
Organisation: text
Intended use: text
I agree to use this model under the Hebrish terms of use (non-commercial unless I hold a commercial licence): checkbox
I agree to receive Hebrish model updates by email (at most once a month; unsubscribe any time): checkbox
---
# Sabich v4.0
<!-- hebrish-status:start -->
> [!TIP]
> **New to Hebrish?** The [Hebrish Index](https://huggingface.co/spaces/Hebrish/Index) explains the project and which model to use: Sabich Rolling for dictation, Falafel for voice commands, Shakshuka for live streaming.
<!-- hebrish-status:end -->
**Sabich** is the speech recognition model in the [Hebrish](https://hebrish.app) apps: English dictation with Hebrew (and
some Yiddish) words dropped in, written in Latin letters, plus tech vocabulary, numbers and Israeli names. Fine-tuned from
`nvidia/parakeet-tdt-0.6b-v2`. This repository is the line the apps use and keep updating; the older
[`Hebrish/Sabich`](https://huggingface.co/Hebrish/Sabich) (v1.0 to v2.4) is no longer updated.
## How the training data is recorded (video)
[![How I Record Training Data for My Own Speech-to-Text Model](https://i.ytimg.com/vi/RWTDqKSKOQE/hqdefault.jpg)](https://www.youtube.com/watch?v=RWTDqKSKOQE)
Daniel on camera: how the takes this model is trained on are recorded, with the capture app
([watch on YouTube](https://www.youtube.com/watch?v=RWTDqKSKOQE); more at [hebrish.app/updates/videos](https://hebrish.app/updates/videos)).
## v4.0 (2026-11-10): a fresh baseline, trained from scratch
Every release in this repository before it (v2.5 to v3.4, from 2026-11-02) was an incremental update that continued from the previous weights. v4.0 is not:
it was trained **from scratch** from stock `nvidia/parakeet-tdt-0.6b-v2` on the whole, freshly standardised dataset, and
it is the new baseline that later updates continue from.
- **Dataset standardisation first (2026-11-09).** One set of transcript conventions was settled and applied to every
published take: `7 PM` and `4:30 AM`; units abbreviated with a space (`150 km`); digits for 10 and above and words for
one to nine in ordinary counts (quoted texts such as Tanach and Talmud exempt); times in digits; one spelling for each
same-sounding term (`kupat cholim`, `Be'er Sheva`); different words kept apart (Jaffa / Yafo, Ram / RAM); names never
respelled; canonical Hebrew romanisations and proper-noun casing; shekel amounts as NIS. Changed rows keep their old
text in `text_before_normalization`. Each topic also became its own Hebrish dataset in one standard format.
The aim is consistent formatting in the output.
- **Training compute: [Nebius](https://nebius.com)** (Nebius AI Cloud, Serverless AI jobs, one NVIDIA H100, eu-north1).
The first Hebrish model trained there; earlier releases were trained on Modal.
- **Recipe:** 20 epochs, learning rate 3e-5, batch 16, speed perturbation 0.9 to 1.1x, public-English replay
(LibriSpeech-PC). 192 minutes of training. The released weights are the best epoch (17).
- **Held-out validation WER: 4.82%** on 1,857 clips never trained on. The validation set grows with the data, so this is
not comparable with earlier releases' numbers, and it is not a benchmark.
### Dataset at training time
Master dataset revision `b87b8df` (2026-11-10): **29,099 clips, 2,754 minutes (about 45.9 hours)** of unique audio, recorded
2025-11-17 to 2026-11-10. Priority-word clips are repeated three times, so the training set was 32,619 rows (3,029 minutes).
| Source | Minutes |
| --- | ---: |
| Daniel's own recordings (phone) | 2,536 |
| LibriSpeech-PC (public English replay) | 191 |
| EHMS import | 27 |
Code-switching sentences: 1,106 minutes; everything else (tech, names, general English, tasks): 1,648 minutes.
| Category (first tag) | Clips | Minutes |
| --- | ---: | ---: |
| tech | 7,007 | 668 |
| message | 3,724 | 394 |
| code_switch | 2,660 | 225 |
| names | 2,512 | 187 |
| voice-commands | 2,418 | 87 |
| email | 1,607 | 199 |
| general | 1,089 | 105 |
| librispeech | 900 | 191 |
| israeli-streets | 577 | 54 |
| dates-times | 406 | 29 |
| brand-names | 362 | 40 |
| whispered | 292 | 27 |
| tanach | 291 | 36 |
| numerics | 282 | 19 |
| irish-english | 242 | 22 |
| other tasks (architecture, October 7, rabbis, web and email addresses, Talmud, rabbinic works, stories, ...) | 4,730 | 471 |
## Earlier versions
- **v2.5 to v3.4** (2026-11-02 to 2026-11-10): incremental updates, each continuing from the previous weights on the
cumulative recordings; the first started from Sabich v2.4.
- **v1.0 to v2.4**: [`Hebrish/Sabich`](https://huggingface.co/Hebrish/Sabich).
## Files
- `Sabich-Rolling.nemo`: the NeMo model (`ASRModel.restore_from("Sabich-Rolling.nemo")`).
- [`Hebrish/Sabich-Rolling-sherpa-onnx-int8`](https://huggingface.co/Hebrish/Sabich-Rolling-sherpa-onnx-int8): the int8 sherpa-onnx build the apps use.
- `training/<run>/config.json` and `result.json`: the exact settings and result of each run.
Contact: access@hebrish.app