Dictation_Dataset / README.md
manasdhir04's picture
Simplify card intro and link AMEND scorer
ab671bb verified
|
Raw History Blame Contribute Delete
3.52 kB
metadata
dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: id
      dtype: string
    - name: spoken_text
      dtype: string
    - name: intended_text
      dtype: string
    - name: domain
      dtype: string
    - name: audio_length_s
      dtype: float64
    - name: speaker_id
      dtype: int64
    - name: gender
      dtype: string
    - name: age
      dtype: int64
    - name: current_city
      dtype: string
  splits:
    - name: test
      num_examples: 151
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
pretty_name: Voice Arena Dictation Dataset (public test)
language:
  - en
license: cc-by-nc-4.0
task_categories:
  - automatic-speech-recognition
size_categories:
  - n<1K
tags:
  - speech
  - asr
  - benchmark
  - dictation
  - english
  - en-US
  - inverse-text-normalization
  - disfluency

Voice Arena Dictation Dataset (public test)

A US English dictation test set: people speaking the kind of text they would otherwise type — emails, chat messages, notes, code requests and prompts for AI assistants. Every clip carries two references:

  • spoken_text — what was actually said, verbatim, including self-corrections ("seven thousand, wait no, four thousand"), restarts and hedges.
  • intended_text — what should land on the page: corrections applied, numbers, dates, currency, emails and URLs written out in their final form.

This is the public split of the Voice Arena dictation eval.

At a glance

Clips 151
Duration 0.73 h
Speakers 71
Clip length 17.5 s mean / 12.0 s median (3.1 s – 90.3 s)
Words (spoken) 6,922
Female / male / other clips 124 / 25 / 2
Speaker age 23 – 72 (median 47)
US states (current residence) 28
Clips where spoken ≠ intended 121

Domains

domain Clips
code 26
personal_messaging 25
work_chat 24
ai_prompt 22
docs_notes 22
email 22
long_form 10

Speech phenomena

Clips
Named entities (people, dates, amounts, emails/URLs, …) 74
Questions 43
Self-corrections 27
Spoken lists 26
Restarts 23
Hedges 19

Loading

from datasets import load_dataset

ds = load_dataset("VoiceArena/Dictation_Dataset", split="test")
print(ds[0]["spoken_text"])
print(ds[0]["intended_text"])

Audio is mono FLAC at 44.1 kHz.

Data fields

Field Type Description
audio Audio The speech clip (FLAC).
id string Unique clip identifier.
spoken_text string Verbatim reference of what was said, including self-corrections, restarts and hedges.
intended_text string The text the speaker meant to produce: corrections resolved and written-form formatting applied.
domain string Writing context: email, work_chat, personal_messaging, docs_notes, code, ai_prompt or long_form.
audio_length_s float Clip duration in seconds.
speaker_id int Anonymised speaker identifier, consistent across all clips from the same speaker.
gender string Speaker's self-reported gender: female, male or other.
age int Speaker age in years.
current_city string Speaker's city of residence as "city, state, county".

Scoring

To score your own model on this set, use the official scorer: