Proofed

Proofed is a project to build an integrated automatic speech recognition and post-processing model that converts speech directly into clean, proofed text. It starts from a stock NVIDIA Parakeet TDT 0.6B v2 model, fine-tuned only on cleanup tasks so that transcripts come out already edited: filler words such as "um" are removed, repeats and false starts are dropped, numbers are written as digits, self-corrections are applied, remarks addressed to someone else in the room are omitted, and whispered speech is handled. The goal is a single small model performing one pass, without a language model applied afterwards. Domain-specific vocabulary is explicitly out of scope and is planned as a later layer.

The project is at an early stage. Phase 1 research, covering feasibility, prior art, base models, and training data, is complete and concluded the approach is viable, but no model has been trained yet. The repository contains a task taxonomy classifying each post-processing task by operation type and tier, a survey of paragraph-token and newline support across NVIDIA Parakeet, Canary, and Nemotron speech models, and recording targets for the first fine-tune. A public training dataset of 600 takes, roughly 75 minutes of audio, is published on Hugging Face under the CC-BY-4.0 licence.

The project may be of interest to developers building dictation, meeting transcription, or voice note tools who want readable transcripts without a separate editing step or an additional LLM pass. It also documents a worked approach to adapting a general-purpose ASR model for a specific output style, including how the training data and task categories were defined.