MimicScribe
An on-device speech transcriber built on a custom-trained Parakeet Ultra model. The custom training fixes the model family's issues with digits by having the model output them in spoken form, and the harness it sits in is being improved to better merge 15-second audio windows for more accurate transcription.
What it addresses
The Parakeet model family has accuracy issues transcribing digits, and the surrounding harness merges 15s windows poorly.
Shared over time
Each entry reflects what was shared on that date. Details may have changed since then.
MimicScribe
An on-device speech transcriber built on a custom-trained Parakeet Ultra model. The custom training fixes the model family's issues with digits by having the model output them in spoken form, and the harness it sits in is being improved to better merge 15-second audio windows for more accurate transcription.
Problem: The Parakeet model family has accuracy issues transcribing digits, and the surrounding harness merges 15s windows poorly.
Posted by @marshalla · Original comment ↗
“I'm trying to build "the most accurate on-device transcriber". I'm doing a custom train of Parakeet Ultra that fixes the model families' issues with digits by having the model output them in spoken form. I'm also trying to get the harness it sits in to better merge the 15s windows. https://mimicscribe.app/”