ductai199x/dictation-speech

Dictation Speech

103,486 clips (563 hours) of dictation: people's kinds of text, spoken the way people dictate (fillers, false starts, self-corrections, editing commands, spoken punctuation, lists, grammar slips), each with what was said word for word and the text the speaker meant to type. 16 kHz mono, stored losslessly as FLAC. All speech is synthetic (Kokoro-82M and Chatterbox).

It is the speech side of Utter's data:

• the corpus* collections are the audio behind ductai199x/dictation-to-text, one clip per text pair (join rendition_id to its id), each in its pair's split;
• the other collections are the eval sets Utter's models are measured on, all test, including voices (Chatterbox) no training clip uses.

Dataset structure

split	clips	hours
train	79,246	435
test	24,240	128

Configs: all (the default: every collection, train and test) and one per collection:

collection	train	test	hours	categories
corpus	15,857	1,992	122.1	command, correction, disfluent, explicit, natural, restate
corpus-v2	24,419	3,081	150.7	command, correction, disfluent, explicit, natural, restate
corpus-messages	11,080	1,398	66.6	amend, command, correction, disfluent, explicit, grammar, natural, restate
corpus-amend	7,499	947	46.2	amend
corpus-grammar	7,759	1,018	40.1	grammar
corpus-grammar2	9,102	1,113	46.0	grammar
corpus-lists	3,530	446	17.5	list
kokoro		4,642	24.7	command, correction, disfluent, explicit, natural
kokoro-amend		1,166	6.4	amend
kokoro-directed		292	1.4	disfluent, natural
kokoro-grammar		676	3.2	grammar
kokoro-lists		803	3.5	list
kokoro-real		937	4.1	amend, command, correction, disfluent, explicit, grammar, natural
kokoro-restate		563	4.5	restate
chatterbox		4,229	21.8	command, correction, disfluent, explicit, natural
chatterbox-real		937	3.8	amend, command, correction, disfluent, explicit, grammar, natural

The same rendition can be in more than one test collection (spoken by both engines, or in a focused set and the main one); id is unique, rendition_id is shared. No rendition and no passage is in both splits.

column	type	meaning
id	string	<collection>/<rendition id>, unique
rendition_id	string	<passage id>-<category>; the id of the same row in dictation-to-text
audio	Audio (16 kHz)	the clip, FLAC
duration_s	float32	its length
language	string	en
transcript	string	exactly what was said, as spoken (numbers and symbols in words)
final_text	string	the text the speaker meant to type
category	string	how it was spoken (below)
domain	string	what the passage is
annotations	list of strings	what the rendition contains, as key=value (e.g. fillers=2, false_start=1, errors=VERB:SVA, op=value)
tts_engine	string	kokoro-82m (98,320), chatterbox-0.1.7 (5,166)
tts_voice	string	the voice (Kokoro's voice name, or a numbered donated reference voice for Chatterbox)
tts_settings	list of strings	speed= (Kokoro) or exaggeration= (Chatterbox); audio=regenerated on clips spoken again (below)
collection	string	the set it belongs to
source	string	where the passage came from
source_url	string	the passage's page (for passages Utter's tools wrote, the script that wrote them)
license	string	the passage's license

Categories:

category	spoken	clips
grammar	with the grammar slips of real speakers, to be corrected	21,131
disfluent	with fillers, repeats and false starts	11,901
command	with a spoken editing command ("scratch that")	11,826
natural	the passage said plainly	11,669
amend	with a later edit to something said earlier	11,533
correction	with a spoken self-correction ("three, no, four")	11,522
explicit	with spoken punctuation ("comma", "new paragraph")	11,408
restate	said again in other words partway through	7,717
list	a list, spoken item by item	4,779

Sources and licensing

source	clips
common-pile/stackexchange	36,354
synthetic	19,690
common-pile/arxiv_abstracts	17,740
common-pile/ubuntu_irc	15,133
wikimedia/wikipedia	14,569

license	clips
cc-by-sa-4.0	42,611
cc-by-sa-3.0	20,888
cc0-1.0	17,740
public-domain	15,133
cc-by-sa-2.5	7,114

The dataset is CC BY-SA 4.0, as most of its text requires; each row keeps its passage's license and source URL (the page with the passage and its author).

The speech: Kokoro-82M is Apache-2.0, and its output is unrestricted; Chatterbox (MIT) cloned voices from kyutai/tts-voices (voice-donations, CC0: recordings people donated to the public domain).

Dataset creation

1. Passages (Common Pile, Wikipedia, and everyday messages written for the dataset), rewritten as spoken renditions by a language model (Gemma 4 26B) in the categories above.
2. Spoken by Kokoro-82M (twelve English voices, rate 0.9-1.1, chosen per rendition by a hash of its id) or Chatterbox (voice-cloned from Kyutai's CC0 donated voices, exaggeration varied), resampled to 16 kHz, with 0.25 s of silence around.
3. 6,135 corpus clips (audio=regenerated) are a second synthesis with the same text, voice and speed: the same content and length to within a few hundredths of a second, not the same samples as the audio dictation-to-text's transcripts were made from.
4. Final texts: the passage, with grammar corrected for the grammar category and mechanical punctuation slips fixed; the same as dictation-to-text's.

Personal and sensitive information

All speech is synthetic; no recording of a real person is included. The Chatterbox voices imitate recordings people donated to the public domain (kyutai/tts-voices); they are numbered here, not named, and must not be used to identify or impersonate the donors. Passage text from public posts can contain the usernames their writers published.

Bias, risks and limitations

• English only; synthetic voices (twelve Kokoro voices and a small set of donated ones), so real accents, rooms and microphones are not represented. Use it for dictation formatting and cleanup, and as an ASR stress test, not as a substitute for real speech.
• Technical forums are overrepresented next to everyday writing.

Use

from datasets import load_dataset

test = load_dataset("ductai199x/dictation-speech", split="test")             # every eval collection
amend = load_dataset("ductai199x/dictation-speech", "kokoro-amend", split="test")
row = test[0]; row["audio"]["array"], row["transcript"], row["final_text"]

Citation

@misc{nguyen2026dictationspeech,
  author       = {Tai Duc Nguyen},
  title        = {Dictation Speech},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/ductai199x/dictation-speech}}
}