Hugging Face's logo

Datasets:
Kratos-AI
/
KAI_speech-recognition-data

Modalities:
Audio
Text
Formats:
soundfolder
Size:
< 1K
Libraries:
Datasets
Dataset card Data Studio Files
xet
Community
2
KAI_speech-recognition-data / README.md
KGeN
Update README.md
c2e2247 verified
preview code
|
raw
history blame
3.59 kB

Audio Speech Recognition Dataset

Dataset Card

Dataset Summary

This dataset comprises high-quality audio recordings of a standardized 50-word paragraph, designed for training and evaluating automatic speech recognition (ASR) models. Each recording captures the text: "AI is transforming our world, from voice assistants to self-driving cars. Good AI needs quality data. Your voice helps train speech recognition. Speak clearly, naturally, without noise. Be yourself. Thanks for improving AI!" The dataset supports the development of robust speech recognition systems by providing clear, unedited voice samples contributed by individuals worldwide.

Supported Tasks and Leaderboards

  • Task Categories: Automatic Speech Recognition (ASR)
  • Supported Tasks:
    • Speech-to-text transcription
    • Audio signal processing
    • Voice synthesis (secondary use)
    • Natural language processing (NLP) applications (secondary use)

Languages

  • Primary Language: English (en)
  • Future Potential: Multilingual recordings may be added in future versions.

Dataset Structure

Data Instances

Each instance consists of an audio file and corresponding metadata:

  • Audio: Stored in the audio/ directory with filenames like audio1.mp3, audio2.mp3, etc.
  • Formats: MP3, WAV, or M4A
  • Duration: 20–25 seconds per recording

Data Fields

The metadata.csv file includes:

  • file_name: Path to the audio file (e.g., audio.mp3)
  • text: e.x., The transcribed 50-word paragraph: "AI is transforming our world, from voice assistants to self-driving cars. Good AI needs quality data. Your voice helps train speech recognition. Speak clearly, naturally, without noise. Be yourself. Thanks for improving AI!"

Dataset Creation

Curation Rationale

This dataset was created to provide high-quality, standardized audio data for training and evaluating ASR models. The consistent 50-word paragraph ensures uniformity while allowing diversity in accents and speaking styles, making it ideal for improving speech-to-text accuracy.

Source Data

  • Contributors: Individuals worldwide who recorded the standardized paragraph.
  • Recording Guidelines: Contributors record in a quiet environment, using their natural speaking style, without background noise, music, or other speakers. Recordings are original, unedited, and 20–25 seconds long.

Annotations

  • Annotation Process: The metadata.csv file includes manual transcriptions (the fixed 50-word paragraph) and metadata such as file name, duration, and license.
  • Annotators: Dataset maintainers and contributors.

Considerations for Using the Data

Social Impact of Dataset

This dataset aims to advance AI speech recognition by providing diverse, high-quality voice samples, potentially improving accessibility and usability of voice-driven technologies. However, care should be taken to ensure fair representation across demographics to avoid bias in trained models.

Discussion of Biases

  • Current Limitation: With fewer than 1,000 samples, the dataset may lack sufficient diversity in age, gender, and accents.
  • Mitigation: Encouraging contributions from varied demographics to enhance representativeness.

Other Known Limitations

  • Size: The dataset is small (<1,000 samples), limiting its robustness for large-scale ASR training.
  • Language: Currently English-only; multilingual expansion is planned.
  • Out-of-Scope Use: Not designed for real-time speech applications or commercial use without adhering to the CC BY 4.0 license.