nyalpatel/condensed_librispeech_asr

Condensed LibriSpeech ASR

This dataset is a condensed version of the LibriSpeech ASR dataset, created by subsampling approximately 10% of the original data from each split. It is intended for quick experimentation, prototyping, and debugging when working with Automatic Speech Recognition (ASR) tasks.

Dataset Details

• Original Dataset: LibriSpeech ASR
• Condensation Ratio: Approximately 10% of the full dataset
• Splits Included:
  • train.clean.100
  • train.clean.360
  • train.other.500
  • validation.clean
  • validation.other
  • test.clean
  • test.other

For each split, a default number of examples was extracted:

• Training Splits: 1,000 examples each
• Validation/Test Splits: 100 examples each

Data Format

Each sample in the dataset contains the following fields:

• file:
  A path to the original audio file (FLAC format).

• audio:
  A dictionary containing:

  • path: Path to the audio file.
  • array: The decoded audio waveform as a NumPy array.
  • sampling_rate: The sampling rate (typically 16 kHz).

• text:
  The transcription corresponding to the audio.

• id:
  A unique identifier for the sample.

• speaker_id:
  A unique identifier for the speaker.

• chapter_id:
  An identifier corresponding to the audiobook chapter.

How Was This Dataset Created?

The condensed dataset was generated by streaming the full LibriSpeech ASR dataset using the Hugging Face Datasets library and selecting approximately 10% of each split. This process preserves the original structure and fields, enabling seamless use with models and workflows designed for LibriSpeech.

Usage Example

Below is a Python snippet to load and inspect the dataset:

from datasets import load_dataset

# Load the condensed dataset from the Hugging Face Hub
dataset = load_dataset("nyalpatel/condensed_librispeech_asr")

# Access a specific split (e.g., test.clean)
test_dataset = dataset["test.clean"]

# Display the first example in the test set
print(test_dataset[0])