Models · NVIDIA · Parakeet TDT 0.6B v2
Fritter
Daniel's English dictation model, for speech with no Hebrew in it. Not part of the Hebrish family: a personal fine-tune, listed here because it comes from the same work.
At a glance
| Base model | nvidia/parakeet-tdt-0.6b-v2 (NVIDIA) |
| Current version | Fritter v1.2, released 2026-11-05 |
| Parameters | 0.6B (base model's size) |
| Evaluation | v1.2 not benchmarked yet. v1.0 on its own test clips (not comparable with the Hebrish test sets): 4.01% tech WER (stock 7.53%), 4.73% Hebrew-free English WER (stock 7.46%) Results on the model card · Browser demo (runs v1.0) |
| Access | Direct download: no request or sign-in needed. Download on Hugging Face |
| Download | every format and quant (Hugging Face) |
| Licence | CC-BY-NC-4.0 |
| Used by | transcribe.cpp (GGUF), sherpa-onnx and NeMo, by hand |
| DOIs | 10.57967/hf/11384 |
| Full model card | results, recipe, limitations |
Formats and quants
| Repository | Format and files | Version | For | DOI |
|---|---|---|---|---|
danielrosehill/Fritter main | .nemo, GGUF, ONNX, sherpa-onnx int8 .nemo 2.47 GB; GGUF Q8_0 730 MB, Q4_K_M 476 MB, F16 1.24 GB; ONNX fp32 and int8; sherpa-onnx int8 encoder 652 MB · direct download · files | Fritter v1.2 | Desktop and phone, in any app that loads these formats | 10.57967/hf/11384 |
danielrosehill/Fritter-Nemotron-0.6B Nemotron base | .nemo (NeMo) .nemo, GGUF F16, training configs · needs approval · files | see the model card | NVIDIA NeMo; Fritter on the Nemotron streaming base | none yet |
danielrosehill/Fritter-Nemotron-0.6B-sherpa-onnx-int8 Nemotron base | sherpa-onnx int8 encoder, decoder, joiner int8 · needs approval · files | see the model card | sherpa-onnx apps | none yet |
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 2026-11-06.
Versions
| Version | Released | Run | Where | Notes |
|---|---|---|---|---|
| Fritter v1.2 current | 2026-11-05 | r30-general | Fritter | Not benchmarked yet |
| Fritter v1.1 | 2026-11-03 | r25-general | Fritter | About 2.2 times v1.0's audio (694.8 min of own takes). Not benchmarked yet |
| Fritter v1.0 | 2026-10-27 | r11-general | Fritter | 391.0 min, 3,848 clips |
About it
Hebrish models learn the Hebrew words in English speech. Fritter does the opposite job: it leaves code-switching out and concentrates on the English Daniel dictates all day: model and product names, command-line and cloud terms, brand names, acronyms, numbers and currency in digits, Irish-English phrasing, and everyday emails and messages. It is a full fine-tune of nvidia/parakeet-tdt-0.6b-v2.
If you say Hebrew words, use Sabich. Fritter was not trained on code-switching, so Hebrew words come out the way stock Parakeet writes them. It isn't offered in the Hebrish apps.
Fritter v1.1 has not been benchmarked yet. These are Fritter v1.0's results from the model card, on its own test clips (never trained on):
| Word error rate | Stock Parakeet | Fritter v1.0 |
|---|---|---|
| Hebrew-free English, 319 clips | 7.46% | 4.73% |
| Tech sentences | 7.53% | 4.01% |
| Own English, 364 clips | 7.23% | 4.85% |
| LibriSpeech, 200 utterances | 1.23% | 1.47% |
These test clips are not a Hebrish test set, so the numbers do not compare with the ones on /benchmarks. Fritter is adapted to one speaker's voice, accent and phone, so other people will see less benefit, and the test sets are small: differences under about one point are within noise.
Try it
Fritter-Demo runs the int8 ONNX build in your browser, on your own device. The demo still runs Fritter v1.0.
The GGUF builds run on transcribe.cpp; Q8_0 is the recommended one.
Cite
Cite Fritter with its DOI, 10.57967/hf/11384 (danielrosehill/Fritter). To cite one format, use that repository's DOI from the table above.
@misc{rosehill_fritter_2026,
author = {Daniel Rosehill},
title = {Fritter},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/danielrosehill/Fritter},
doi = {10.57967/hf/11384}
}Access
Open: no access request needed. Download Fritter from Hugging Face and use it under the CC-BY-NC-4.0 licence.
Access by repository:
- Direct download:
danielrosehill/Fritter - Needs approval:
danielrosehill/Fritter-Nemotron-0.6B,danielrosehill/Fritter-Nemotron-0.6B-sherpa-onnx-int8