---
license: cc-by-4.0
language:
- en
task_categories:
- automatic-speech-recognition
tags:
- asr
- disfluency
- inverse-text-normalization
- post-processing
- parakeet
- proofed
pretty_name: Proofed ASR Cleanup Training Data
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train/**
  - split: test
    path: data/test/**
---

# 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](https://github.com/danielrosehill/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, single speaker. The per-task
takes are still being recorded (targets: [recording-targets.md](https://github.com/danielrosehill/Proofed/blob/main/docs/recording-targets.md));
this repo is rebuilt as they land. Task definitions: [taxonomy](https://github.com/danielrosehill/Proofed/blob/main/docs/taxonomy.md).

## Contents

| Task | Train takes | Train min | Test takes |
|---|---|---|---|
| replay (plain English, against forgetting) | 344 | 34.3 | 19 |
| currency | 116 | 11.1 | 0 |
| disfluency (filled pauses) | 51 | 12.4 | 4 |
| numerics | 37 | 2.7 | 4 |
| repetitions | 18 | 3.6 | 1 |
| false-starts | 3 | 0.3 | 0 |
| paragraphs | 3 | 0.7 | 0 |
| whispered | 3 | 0.4 | 1 |
| asides | 2 | 0.3 | 0 |

Self-corrections have no takes yet.

## Columns

- `audio` (from `file_name`): 16 kHz mono WAV, recorded on an Android phone.
- `text`: **the clean training target** (what Proofed should write).
- `verbatim_text` / `spoken_text`: what was actually said, where it differs (disfluent, repeated,
  self-corrected takes); `aside_text`, `aside_position`: the aside that must be dropped;
  `paragraph_text`: the target with paragraph breaks.
- `task`, `tasks`, `categories`: task labels (ids as in the taxonomy).
- `duration_s`, `device`, `microphone`, `environment`, `sample_rate`, `recorded_at`, `mode`,
  `prompt_id`, `prompt_source`, `prompt_kind`.
- `source_dataset`, `source_revision`: provenance of each row.

## Sources

Built from [`danielrosehill/Read-Aloud-Voice-Sentences`](https://huggingface.co/datasets/danielrosehill/Read-Aloud-Voice-Sentences)
@ `c544f891` and the task subsets `Hebrish/Disfluent-Read-Sentences`, `Hebrish/Repeated-Word-Sentences`,
`Hebrish/Currency-Sentences`, `Hebrish/Spoken-Aside-Sentences`, `Hebrish/Self-Correction-Sentences`
(revisions in `build_summary.json`). Train/test splits follow the source datasets. Code-switching,
tech and other domain-vocabulary takes are excluded by design. Read, acted speech: disfluencies
were performed from prompts, not spontaneous.

Licence: CC-BY-4.0. Non-commercial research project; public corpora used later in training
(DisfluencySpeech, AMI, ICSI, LibriSpeech-PC) are not redistributed here.