Why it exists
General speech models are excellent at everyday English and weak at jargon. Say "pull the Qwen3 weights with Ollama and test them in vLLM" and the names can easily come back as ordinary English words. For someone who dictates all day to AI tools and terminals, those misses add up to constant correction.
Vocabulary
Thousands of tech terms, each read in several natural sentences, so the model learns the spelling and casing as well as the sound.
Formatting
Numbers, versions and prices come out as digits: "GPT-5.1", "$1,850", "20%", not words.
One voice
Trained on one Irish English speaker. It's a personal model, published so others can see how far a few hours of targeted audio goes.
Kept current
New model and tool names are added every week and the model is updated from its last weights, so the list keeps growing.
Large downloadThe model is about 660 MB. Your browser downloads it once, then keeps it. Best on a desktop or laptop in Chrome, Edge or Firefox; once loaded it uses about 2.5 GB of memory.
Fritter is adapted to one speaker (an Irish English speaker). Other voices will see less of a gain over stock Parakeet.
Open the full page to try it
Here, inside Hugging Face's frame, your browser won't let the model use more than one processor core, and it would freeze while it starts. The full page runs it properly.
1. Load the model
Checking whether the model is already saved in this browser…
2. Say something
Press Record, read a sentence below (or your own), then press Stop. Up to 30 seconds.
What Fritter heard
Sentences to read aloud
Product names and terms should come out spelt and cased as shown. Numbers come out as digits.
- Can you check the GitHub Actions logs before we deploy to Vercel?
- Run kubectl get pods and send me anything stuck in CrashLoopBackOff.
- Ask Claude to write the PostgreSQL migration, then test it on Supabase.
- The RTX 4090 came to $1,850, about 20% over budget.
Results
Measured at v1.0 (v1.1 to v1.3 are incremental updates on more audio and have not been re-benchmarked). On held-out test recordings from the same speaker (term-disjoint from training), against stock nvidia/parakeet-tdt-0.6b-v2. WER is normalised.
| Stock v2 | Fritter v1.0 | |
|---|---|---|
| Tech sentences, WER (213 clips) | 7.53% | 4.01% |
| All English, WER (364 clips) | 7.23% | 4.85% |
| Target term recall (447 occurrences) | 65.1% | 79.6% |
| Numbers recall (62 tokens) | 33.9% | 50.0% |
| LibriSpeech test-clean, WER (200) | 1.23% | 1.47% |
Full table and method: the model card.
How this page works
It runs the int8 ONNX build of Fritter v1.3, downloaded from Hugging Face, with parakeet.js and ONNX Runtime Web on your processor. The files are kept in your browser's storage; Remove the saved model deletes them.
Follow along and related projects
Releases, data growth and a dated log are on the public tracker, github.com/Fritter-ASR/Fritter. Other open speech-tech work: