Fritter, my English dictation fine-tune for technical vocabulary, is now at v1.3. Two other things changed today: the browser demo has a new front page, and the project has a public tracker repo.
v1.3
v1.3 is another incremental update. It continues from v1.2's weights on 184.7 minutes of recordings I made since then (about 1,800 takes), with a replay sample mixed in so it doesn't lose what it already knew. Across all the releases that comes to about 18 hours of my own voice reading tech sentences aloud.
I haven't re-benchmarked it. The only measured numbers are still 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. (Each update since has been trained on more data, but "more data" is not a measurement, so I won't claim a new figure.) It is still trained on one speaker, me, so other voices may gain less.
A demo you can look at
A speech model is a fairly dry thing to show off. You can't see what it knows until you talk to it, and most people landing on a model page won't. So the demo Space now opens with a slow-moving cloud of the actual terms Fritter was trained on, taken from the training takes: 320 of them, coloured by group and sized by how many times I recorded each one.
Hugging Face Space danielrosehill/Fritter-Demo Parakeet v2 fine-tune for tech dictation, in your browser StaticLiveCC-BY-NC-4.0Updated Nov 2026 Open the SpaceThe groups are AI and models (MCP, Claude Code, Qwen, Ollama), speech and audio (STT, Parakeet, eSpeak, disfluency), dev and data (Typst, Infisical, Jupyter, arXiv), Linux and self-hosting (Konsole, Kdenlive, Debian, rclone), smart home (OPNsense, Modbus, Reolink, Z-Wave JS) and hardware (ThinkPad, VRAM, NVMe, 8BitDo). Hover over or tap a word and it shows a real sentence I recorded it in. Below the cloud, the model itself still runs in your browser, on your own device.
The same cloud is now on the model's page on my site, ai.danielrosehill.com/models/fritter.
A public tracker
Fritter's releases were recorded in a private notes repo until now. The new danielrosehill/Fritter repo is the public version: every release with its date, how it was trained and how much audio went in, plus a dated log. It holds no code or weights; those stay on Hugging Face.
GitHub Repository danielrosehill/Fritter Public tracker for Fritter, a Parakeet TDT 0.6B v2 fine-tune for English tech dictation: releases, data, log Updated Nov 2026 View on GitHubLinks
Fritter on Hugging Face: .nemo, GGUF for transcribe.cpp, ONNX and sherpa-onnx int8
Fritter Nemotron, the streaming version (still v1.0, not benchmarked)
Fritter is CC-BY-NC-4.0: free for non-commercial use with attribution. If you try it on a voice that isn't mine, I'd still like to hear how it goes.
