PardisTaghavi/coraal_chunked

CORAAL Atlanta Dataset - Cleaned for ASR

Dataset Description

This dataset contains cleaned and segmented audio from the Corpus of Regional African American Language (CORAAL) Atlanta subset (v. 2020.05), specifically prepared for Automatic Speech Recognition (ASR) evaluation.

Dataset Summary

• Source: CORAAL (Corpus of Regional African American Language) - Atlanta subset
• Version: CORAAL:ATL v. 2020.05
• Language: African American English (AAE)
• Task: Automatic Speech Recognition
• Audio Duration: 1-20 seconds per segment
• Sampling Rate: 16kHz
• Format: WAV files with corresponding transcriptions
• Total Samples: ~1,500+ audio segments
• Speakers: Multiple AAE speakers from Atlanta

Data Processing

The original CORAAL audio files (40-60 minutes each) have been:

1. Segmented into individual utterances using provided timestamps

2. Cleaned to remove:

   • Interviewer speech (keeping only main speakers)
   • Markup annotations like [Bye.], [/Oh man/.]
   • Pause markers like (pause 0.28)
   • Anonymized names like /RD-NAME-2/
   • Very short segments (< 2 words)
   • Very long segments (> 20 seconds)

3. Normalized transcriptions:

   • Converted to lowercase
   • Removed punctuation
   • Cleaned extra whitespace

Dataset Structure

{
  "id": "ATL_se0_ag1_f_01_1_seg_000001",
  "audio": {
    "path": "ATL_se0_ag1_f_01_1_seg_000001.wav",
    "array": array([-0.00048828, -0.00018311, ...]),
    "sampling_rate": 16000
  },
  "transcription": "they talking about dont send him to his daddy",
  "speaker_id": "ag1",
  "session_id": "ATL_se0_ag1_f_01"
}

Data Splits

• Train: 80% of speakers (speaker-independent split)
• Test: 20% of speakers (no speaker overlap with train)

This ensures proper evaluation for speaker-independent ASR systems.

Speaker Information

• Speakers: Various African American English speakers from Atlanta
• Demographics: Mixed age groups and genders
• Split: Speaker-independent (no speaker overlap between train/test)

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("PardisTaghavi/coraal_chunked")

# Access train split
train_data = dataset["train"]

# Example usage
for example in train_data:
    audio = example["audio"]["array"]
    transcription = example["transcription"]
    speaker_id = example["speaker_id"]
    # Process for your ASR model

Example Training Loop

import torch
from datasets import load_dataset
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC

# Load dataset
dataset = load_dataset("PardisTaghavi/coraal_chunked")

# Load processor and model
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")

# Training loop
for batch in dataset["train"]:
    # Process audio
    inputs = processor(
        batch["audio"]["array"], 
        sampling_rate=16000, 
        return_tensors="pt"
    )
    
    # Process labels
    labels = processor.tokenizer(
        batch["transcription"], 
        return_tensors="pt"
    ).input_ids
    
    # Train your model...

Dataset Statistics

• Total Duration: Several hours of speech
• Average Segment Length: ~5-10 seconds
• Vocabulary: African American English dialectal features
• Quality: High-quality recordings from sociolinguistic interviews

Applications

This dataset is ideal for:

• Training ASR models on African American English
• Evaluating bias in existing ASR systems
• Dialectal speech recognition research
• Sociolinguistic speech technology studies

Citation

If you use this dataset, please cite the original CORAAL corpus:

@misc{coraal,
  title={The Corpus of Regional African American Language},
  author={Kendall, Tyler and Farrington, Charlie},
  year={2023},
  version={2023.06},
  address={Eugene, OR},
  publisher={The Online Resources for African American Language Project},
  doi={10.7264/1ad5-6t35},
  url={https://oraal.uoregon.edu/coraal}
}

License

This dataset is derived from CORAAL and is distributed under the same Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

License Terms:

• ✅ Attribution Required: Must cite original CORAAL corpus
• ✅ NonCommercial Use Only: Research and educational purposes only
• ✅ ShareAlike: Derivatives must use compatible license
• ❌ No Commercial Use: Cannot be used for commercial purposes

Original License: https://creativecommons.org/licenses/by-nc-sa/4.0/

Acknowledgments

• Original CORAAL data: Kendall, Tyler and Charlie Farrington, University of Oregon
• CORAAL Website: https://oraal.uoregon.edu/coraal
• Funding: National Science Foundation (Grant No. BCS-1358724)
• Data cleaning and segmentation: Pardis Taghavi
• Contact: For questions about CORAAL: [email protected]
• Twitter: @CorpusAAL

Contact

For questions about this processed dataset, please open an issue in this repository or contact the dataset creator.

For questions about the original CORAAL corpus, please contact: [email protected]