Streaming ASR Dataset

This dataset is designed for training real-time (streaming) ASR models, with a focus on handling chunk-based audio processing. It contains standardized audio segments from LibriSpeech dev-clean, processed for streaming ASR applications.

Dataset Description

Dataset Summary

Source: LibriSpeech dev-clean
Total chunks: 2,703
Total duration: ~20 hours (1,212.26 seconds)
Unique speakers: 40
Audio format: 16 kHz mono WAV
Language: English
Domain: Audiobooks (clean speech)

Dataset Structure

openwhisper/
├── chunks/          # Audio files (16kHz mono WAV)
├── transcripts/     # Text transcriptions
└── metadata/        # JSON files with detailed information

Data Fields

Each sample consists of:

Audio file (WAV)
- 16 kHz sampling rate
- Mono channel
- 16-bit PCM format
Transcript file (TXT)
- Clean text transcription
- Includes punctuation and casing
- Aligned with audio chunks
Metadata file (JSON)
- speaker_id: Unique speaker identifier
- chunk_id: Unique chunk identifier
- start_time: Start time in original audio
- end_time: End time in original audio
- duration: Chunk duration in seconds
- language: Language code (en)
- noise_conditions: Audio quality label (clean)
- original_file: Source file reference

Data Splits

This dataset contains only the dev-clean portion of LibriSpeech, processed into overlapping chunks suitable for streaming ASR training.

Dataset Creation

Preprocessing

Audio standardization
- Resampling to 16 kHz
- Conversion to mono channel
- Format conversion to WAV
Chunking strategy
- Fixed chunk duration with overlap
- Natural pause boundary detection
- Consistent chunk size for training stability
Transcript processing
- Alignment with audio chunks
- Preservation of punctuation and casing
- Clean text normalization

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("orgho98/openwhisper")

Training Example

# Example code for loading audio and transcript pairs
for sample in dataset:
    audio = sample['audio']
    transcript = sample['text']
    metadata = sample['metadata']
    
    # Process for streaming ASR training
    # ...

License

This dataset is released under the MIT License, following LibriSpeech's licensing terms.

Citation

If you use this dataset, please cite:

@misc{openwhisper2024,
  title={Streaming ASR Dataset},
  author={Automagically AI},
  year={2024},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/orgho98/openwhisper}}
}

Limitations

Limited to clean speech from audiobooks
Single language (English)
May not represent real-world streaming conditions perfectly

Additional Information

Curated by: Automagically AI
License: MIT
Version: 1.0.0

Datasets:

orgho98
/

openwhisper

You need to agree to share your contact information to access this dataset