S1-mini Parakeet Cleanup (LoRA)
Status (2026-10-16): placeholder. No adapter has been trained yet.
A planned LoRA adapter for S1-mini by Superwhisper
(superwhisper/s1-mini, pinned at commit 88f6b15),
tuning it for light-touch cleanup of raw nvidia/parakeet-tdt-0.6b-v2 dictation: fillers, stutters, false starts,
informal contractions and paragraph breaks ([Layers: 1]), plus optional light editorial cleanup such as repeated
points and speech aimed at someone else in the room ([Layers: 1,2]). It keeps wording, numbers and names unchanged.
Project, plan, prompt format and training script: https://github.com/danielrosehill/ASR-Post-Processor-Model
(see docs/fine-tuning.md).
Licence and attribution
This adapter is a derivative of S1-mini by Superwhisper, itself a derivative of Qwen3-0.6B (Alibaba Cloud), both
under Apache 2.0. S1-mini's licence adds one term: any use or derivative must continue to identify it as
"S1-mini" by "Superwhisper". The upstream LICENSE and NOTICE will ship in this repo with the first adapter.
- Downloads last month
- -