Proofed ASR Cleanup Training Data
Proofed ASR Cleanup Training Data Audio paired with clean, proofed target text for training an ASR model to do its own post-processing: no fillers, repeats or false starts, numbers and money as digits, self-corrections applied, asides to other people left out, paragraph breaks, whispered speech. Built for Proofed, a fine-tune of NVIDIA Parakeet TDT 0.6B v2 that writes clean text in one pass with no LLM afterwards. Status (2026-10-24): first snapshot, small. About 75 minutes… See the full description on the dataset page: https://huggingface.co/datasets/danielrosehill/Proofed-ASR-Cleanup-Training-Data.
About This Project
Proofed ASR Cleanup Training Data is a HuggingFace dataset of audio recordings paired with clean, proofed target transcripts. Each row links 16 kHz mono WAV audio, recorded on an Android phone, to a target text that reflects the desired output: fillers, repeats and false starts removed, numbers and currency written as digits, self-corrections applied, side remarks to other people omitted, and paragraph breaks inserted. Where the spoken and target text differ, the dataset also stores the verbatim or spoken text, along with aside and paragraph fields, so that the relationship between what was said and what should be written is explicit. The data is organized by task category, including disfluency, repetitions, currency, numerics, paragraphs, whispered speech and a replay subset of plain English intended to reduce catastrophic forgetting.
The dataset was built to train Proofed, a fine-tune of NVIDIA Parakeet TDT 0.6B v2 that aims to produce clean text in one pass without a subsequent LLM post-processing step. Source material comes from existing Hebrish HuggingFace task subsets and the danielrosehill/Read-Aloud-Voice-Sentences corpus. Disfluencies and other phenomena were performed from prompts rather than captured spontaneously, and the dataset is in an early state, with self-corrections not yet recorded and several categories holding only a few takes. The status is documented as a first snapshot, with the repository planned to be rebuilt as further recordings are added.
The dataset may be of interest to researchers and engineers working on ASR post-processing, disfluency removal, inverse text normalization, or fine-tuning speech models for clean transcription output. It is released under CC-BY-4.0 and is described by its author as a non-commercial research project. Public corpora used in later training stages are not redistributed within this repository.