When I wrote up Fritter on 3 November it was at v1.0, it lived on Hugging Face, and that was more or less the whole story. Three days later that write-up is already behind (which is the trouble with naming a model line and then carrying on recording).
So here is the update. Fritter is now at v1.2, it has a first streaming version, and both can be downloaded directly from a plain URL outside Hugging Face.
What Fritter is
Fritter is an English dictation fine-tune for technical vocabulary, including AI terms: model and product names, CLI and cloud terms, brands, acronyms, and numbers in digits. These are precisely the words I say all day and the ones stock speech recognition trips on.
It is trained on one speaker, namely me. Other voices may gain less (I haven't measured how much less, so I won't guess).
The number I can stand behind is v1.0's: word error rate on tech sentences went from 7.53% on stock Parakeet to 4.01%, on held-out recordings of my own voice. v1.1 and v1.2 haven't been benchmarked, so I'm not claiming anything for them. And it isn't a free lunch: general English came out slightly worse than stock.
How it differs from my code-switching model
I also train a separate English-Hebrew code-switching model, for the English with Hebrew words dropped into it that English speakers in Israel actually talk. Same voice, same habit of reading sentences into a phone, but a different job.
Fritter leaves all of that out. It is trained only on recordings that aren't code-switching, and I also held back the romanised Hebrew that slips past that filter (street names and biblical names, for instance). So a Hebrew word comes out however the stock model would have written it. If you dictate plain English with a lot of jargon, Fritter is the one for you. If your English has Hebrew in it, it isn't.
The streaming version
The offline Fritter transcribes a finished recording. Fritter Nemotron v1.0 writes text as you speak, in 560 ms chunks. It is a full fine-tune of NVIDIA's Nemotron Speech Streaming EN 0.6B on 858.8 minutes of my recordings plus 95.6 minutes of LibriSpeech replay.
It has no benchmark yet. (The run reports a validation figure, but that split only picks the checkpoint, so it tells you nothing about how the model compares with anything.)
Getting it
Both builds are sherpa-onnx int8, about 660 MB each. These URLs always serve the latest version, and each manifest lists the version plus every file's size and sha256:
Fritter v1.2, offline: manifest.json
Fritter Nemotron v1.0, streaming: manifest.json
All the file links on one page: Fritter on Carrot Cake AI
The weights and other builds (GGUF for transcribe.cpp among them) are on Hugging Face, ungated:
Both lines are CC-BY-NC-4.0: free for non-commercial use with attribution. The streaming version's base model is by NVIDIA and licensed by NVIDIA Corporation under the NVIDIA Open Model License, so read the NOTICE file in its Hugging Face repo. This is a personal project, not affiliated with NVIDIA.
If you try either one on a voice that isn't mine, I'd like to hear how it goes. That is the number I don't have.
