milanakdj/augmented-merged-and-shuffel-tibetan-dataset

Augmented Audio Dataset

This dataset contains augmented audio samples with pitch shifting and white noise enhancement for improved model training and robustness.

Dataset Description

This is an augmented version of the original audio dataset, processed with audio augmentation techniques to increase dataset diversity and improve model generalization.

Augmentation Techniques Applied

1. Pitch Shifting: Random pitch shift between -2 to +2 semitones
2. White Noise Addition: Subtle white noise (noise factor: 0.0005-0.002) to simulate real-world recording conditions
3. Mono Conversion: All stereo audio converted to mono for consistency

Dataset Statistics

• Columns: audio, english_text
• Audio Format:
  • Sampling Rate: 16,000 Hz
  • Format: Float32 mono arrays
  • Encoding: Decoded audio arrays (not bytes)

Augmentation Details

Each audio sample has been processed with:

• Pitch Shift Range: -2 to +2 semitones (randomly applied per sample)
• Noise Level: Very subtle (0.0005 to 0.002 amplitude)
• Channel Processing: Converted from stereo to mono where applicable

Use Cases

This augmented dataset is suitable for:

• Speech recognition model training (e.g., Whisper fine-tuning)
• Automatic Speech Recognition (ASR) tasks
• Audio classification
• Robust model development that handles pitch variations and background noise

Data Format

{
    'audio': {
        'array': numpy.ndarray,  # Float32 audio samples
        'sampling_rate': 16000,  # Hz
        'path': str or None
    },
    'english_text': str  # Transcription
}

How to Use

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("your-username/your-augmented-dataset")

# Access audio and text
sample = dataset['train'][0]
audio_array = sample['audio']['array']
sampling_rate = sample['audio']['sampling_rate']
transcription = sample['english_text']

Comparison with Original

• Pitch Variation: Adds ±2 semitones variation to handle different speaker pitches
• Noise Robustness: Subtle noise helps models generalize to non-studio recordings
• Audio Quality: Maintains high quality while adding realistic variations

Citation

If you use this augmented dataset, please cite the original dataset source and mention the augmentation process.

License

[Specify your license - e.g., CC-BY-4.0, MIT, etc.]

Acknowledgments

• Original dataset: [Link to original dataset]
• Augmentation tools: librosa, numpy
• Audio processing: pitch shifting and noise addition

---

Note: All augmentations preserve the original transcriptions. Only audio signals are modified.