fleet-clean-tags / README.md
Lucasgitguy-co's picture
Fleet-Clean release
9640657 verified
|
Raw History Blame Contribute Delete
2.19 kB
metadata
license: cc-by-4.0
language:
  - en
  - de
  - fr
  - es
  - it
  - nl
  - pl
  - ko
task_categories:
  - token-classification
tags:
  - disfluency-removal
  - punctuation-restoration
  - speech-transcripts
  - fleet-2
size_categories:
  - 10K<n<100K

fleet-clean-tags

Word-level cleanup labels on real speech-recognition output: the training data of Fleet-Clean. Every row is a transcript written by Fleet-2 from real audio. Each word has three labels:

Field Values
delete 0 keep, 1 delete
punct mark after the word in the clean text: 0 none, 1 ,, 2 ., 3 ?, 4 !; -100 = unknown
case 0 keep, 1 upper-case the first letter, 2 lower-case it (relative to Fleet-2's casing); -100 = unknown

clean is the text you get by applying the labels, and part says where a row comes from:

  • A: real disfluent English (DisfluencySpeech). Deletions come from its verbatim vs cleaned transcripts.
  • B: fluent speech in 8 languages. Nothing is deleted, and punctuation and case come from the human reference where the words match.
  • C: B rows with synthetic fillers, discourse markers, repeats, restarts and self-corrections inserted (kinds). Only the inserted words are deleted.
Split Rows Words Deleted words
train 28176 649571 33209
validation 1514 35073 1793
test 1585 35892 1918

Splits are by source clip, and a part C row always shares its B row's split. DisfluencySpeech keeps its own splits. No audio and no reference transcripts are included, only Fleet-2's words and the labels.

Sources and licences

This dataset is CC BY 4.0, the most restrictive of these, with attribution to all of the above.