Dictation to Text

89,241 pairs of what a speech recognizer heard someone dictate and the text they meant to type: punctuated, formatted and corrected the way a careful writer would type it, in the speaker's own words. Fillers and false starts go, spoken self-corrections and editing commands are applied, spoken lists become lists, grammar slips are fixed.

It is the training data of Utter's cleanup models (ductai199x/utter). Its speech counterpart, the audio these transcripts came from, is ductai199x/dictation-speech (joined on id = rendition_id).

Dataset structure

The split is by passage (a passage's renditions are all in one split) and balanced: test holds 11.1-11.3% of every category, domain, source, license, length and of the rows with a dictionary or with lists off. No pair appears twice.

Categories:

Sources and licensing

Passages from Common Pile (Stack Exchange, arXiv abstracts, Ubuntu IRC) and Wikipedia, plus everyday messages, emails and notes written for this dataset (synthetic). Every row carries its passage's license:

The dataset as a whole is CC BY-SA 4.0, as most of its text requires (share-alike; CC BY-SA 2.5 and 3.0 allow adaptations under later versions). Rows under CC0 or in the public domain keep that status; text written for the dataset is CC BY-SA 4.0. The passages were adapted (spoken renditions, corrections; see Dataset creation). Attribution: each row's source_url, the page with the passage and its author.

Dataset creation

1. Passages: real text of the kinds people dictate (questions and answers, chat, notes, prose), filtered to English and to lengths people say in one go; everyday messages written by a language model to cover what the public sources lack.
2. Spoken renditions: each passage rewritten the way someone would say it, in one of the categories above, by a language model (Gemma 4 26B).
3. Speech: each rendition spoken by Kokoro-82M (twelve English voices, speaking rate 0.9-1.1).
4. Recognition: Parakeet (the speech model Utter ships) transcribed the audio; its output, errors and habits included, is dictated.
5. Dictionary lines added to about a third of the rows.

Final texts

• The passage itself, as its writer wrote it, with mechanical punctuation slips fixed (?., ,., a space before a comma); grammar-category targets are their passage with its grammar corrected.
• With lists false, a spoken list's target is written as prose.

Personal and sensitive information

Public posts (Stack Exchange, IRC) can contain usernames and the occasional personal detail their writers published. Nothing was added; nothing about the people who use Utter is in this dataset.

Bias, risks and limitations

• English only, written-style text; the spoken side is synthetic speech recognized by one model, so real speakers' accents and recognition errors are only partly represented.
• Technical forums are overrepresented next to everyday writing.
• Targets were produced with a language model's help and checked automatically, not by people, row by row.

Use

from datasets import load_dataset

ds = load_dataset("ductai199x/dictation-to-text")
ds["train"][0]["messages"]   # ready for chat fine-tuning

Citation

@misc{nguyen2026dictation,
  author       = {Tai Duc Nguyen},
  title        = {Dictation to Text},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/ductai199x/dictation-to-text}}
}