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

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

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.

Categories:

Sources and licensing

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}}
}