Repository navigation
Report: single-speaker Parakeet TDT 0.6B v2 fine-tune for tech dictation (Fritter) #16460
danielrosehill
started this conversation in
Public Models
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Hi NeMo team, sharing a community fine-tune of
nvidia/parakeet-tdt-0.6b-v2in case it's useful.Model: danielrosehill/Fritter (public;
.nemoplus GGUF, ONNX and sherpa-onnx int8 builds)Try it in the browser: danielrosehill/Fritter-Demo (int8 ONNX via parakeet.js, runs on-device)
What it is
Single-speaker adaptation for English dictation: tech vocabulary (model and product names, CLI and cloud terms), brand names, acronyms, and numbers written as digits. The speaker has an Irish English accent. Output keeps v2's casing and punctuation.
Method
Results
Held-out takes from the same speaker. Target terms are disjoint from training. WER is normalised.
With 10% replay, general English stays close to stock (LibriSpeech +0.24 points) while the speaker's own tech dictation roughly halves in WER. Long-form dictation came out a little worse, which I haven't investigated yet. The test sets are small and from one speaker, so treat these as indicative.
Thanks for Parakeet. It's a very good base for this kind of personal adaptation.
All reactions