lilgoose777/tibetan-audio-english-sentence-merged-lilgoose

language: - bo license: cc-by-4.0 task_categories: - automatic-speech-recognition - text-to-speech - translation pretty_name: Tibetan Audio-Sentence Merged Dataset size_categories: - 1K<n<10K tags: - tibetan - audio - speech - translation - asr - tts

Tibetan Audio-Sentence Merged Dataset

Dataset Description

This dataset is a comprehensive collection of Tibetan audio recordings paired with their corresponding text transcriptions. It combines three high-quality Tibetan speech datasets to create a larger, more diverse resource for Tibetan language processing tasks.

Dataset Summary

• Language: Tibetan (བོད་ཡིག་)
• Task: Automatic Speech Recognition (ASR), Text-to-Speech (TTS), Translation
• Total Samples: [X samples] (automatically calculated)
• Format: Audio (WAV/MP3) + Tibetan text sentences
• License: CC-BY-4.0

Source Datasets

This dataset merges the following three datasets:

1. lilgoose777/tibetan-speech-english-text-dataset
2. lilgoose777/merged-tibetan-titung-goose
3. Titung/tibetan-to-english-audio-dataset

All source datasets have been standardized to a unified format with two columns: audio and sentence.

Dataset Structure

Data Fields

• audio: Audio file containing Tibetan speech

  • Type: Audio
  • Sample rate: Varies by source (typically 16kHz or 44.1kHz)
  • Format: WAV, MP3, or FLAC

• sentence: Tibetan text transcription

  • Type: string
  • Script: Tibetan Unicode (བོད་ཡིག་)
  • Content: Transcription of the audio content

Data Splits

Currently, this dataset contains a single train split with all samples shuffled randomly.

Split	Samples
train	[X]

Example

from datasets import load_dataset

dataset = load_dataset("your-username/tibetan-audio-sentence-merged")

# Access first sample
sample = dataset['train'][0]
print(sample['sentence'])  # Tibetan text
print(sample['audio'])     # Audio data

Sample Output:

{
  'audio': {
    'path': 'audio_file.wav',
    'array': array([...]),
    'sampling_rate': 16000
  },
  'sentence': 'ཁང་པ་བརྩེགས་པ་རེ་རེར་ཡང་རྫིང་བུ་གཞི་གསེར་གྱི་བྱེ་མ་བདལ་བ།'
}

Dataset Creation

Curation Rationale

The Tibetan language is considered low-resource in terms of digital speech datasets. This merged dataset aims to:

• Increase data availability for Tibetan ASR and TTS research
• Standardize format across multiple sources for easier use
• Provide diverse speakers and contexts by combining multiple datasets
• Support preservation of Tibetan language technology

Source Data

Data Collection

The audio data comes from three independently collected datasets:

• Various Tibetan speakers
• Different recording conditions and qualities
• Mix of read speech and natural conversations
• Both native Tibetan and diaspora speakers

Preprocessing

1. Standardization: All datasets were converted to a unified schema with audio and sentence columns
2. Shuffling: Dataset was randomly shuffled (seed=42) to mix samples from different sources
3. Quality Check: Invalid or incomplete samples were filtered out during processing
4. Format Conversion: Audio files maintained in their original formats

Annotations

Audio files are paired with their Tibetan text transcriptions. The transcriptions are:

• Written in Tibetan Unicode script
• Manually transcribed or verified by native speakers (varies by source)
• Represent the actual spoken content in the audio

Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("your-username/tibetan-audio-sentence-merged")

# Load with streaming (for large datasets)
dataset = load_dataset("your-username/tibetan-audio-sentence-merged", streaming=True)

Example: Automatic Speech Recognition

from datasets import load_dataset
from transformers import pipeline

# Load dataset
dataset = load_dataset("your-username/tibetan-audio-sentence-merged", split="train")

# Load ASR model (example - replace with actual Tibetan model)
asr = pipeline("automatic-speech-recognition", model="your-asr-model")

# Transcribe audio
sample = dataset[0]
transcription = asr(sample['audio'])
print(f"Predicted: {transcription['text']}")
print(f"Reference: {sample['sentence']}")

Example: Text-to-Speech Training

from datasets import load_dataset

dataset = load_dataset("your-username/tibetan-audio-sentence-merged", split="train")

# Prepare for TTS training
for sample in dataset:
    audio = sample['audio']['array']
    text = sample['sentence']
    # Your TTS training code here

Example: Data Analysis

from datasets import load_dataset
import numpy as np

dataset = load_dataset("your-username/tibetan-audio-sentence-merged", split="train")

# Calculate statistics
sentence_lengths = [len(s['sentence']) for s in dataset]
audio_durations = [len(s['audio']['array']) / s['audio']['sampling_rate'] for s in dataset]

print(f"Average sentence length: {np.mean(sentence_lengths):.2f} characters")
print(f"Average audio duration: {np.mean(audio_durations):.2f} seconds")
print(f"Total audio hours: {sum(audio_durations) / 3600:.2f} hours")

Considerations for Using the Data

Social Impact

This dataset contributes to:

• Language Preservation: Supporting Tibetan language technology development
• Accessibility: Enabling voice interfaces for Tibetan speakers
• Education: Facilitating language learning tools
• Cultural Heritage: Preserving spoken Tibetan for future generations

Limitations

• Quality Variation: Audio quality varies across source datasets
• Dialect Coverage: May not represent all Tibetan dialects equally
• Speaker Diversity: Limited information about speaker demographics
• Domain Coverage: May be biased toward certain topics or speaking styles
• Transcription Accuracy: Transcription quality depends on original dataset curation

Ethical Considerations

• Privacy: Ensure audio data does not contain personally identifiable information
• Representation: Dataset may not represent all Tibetan-speaking communities equally
• Cultural Sensitivity: Users should be mindful of cultural context when using this data

Additional Information

Dataset Curators

This merged dataset was created by [Your Name/Organization]. Original datasets were curated by:

• lilgoose777
• Titung

Licensing Information

This dataset is released under CC-BY-4.0 license. Users must:

• Provide attribution to the original creators
• Indicate if changes were made
• Not apply legal terms that restrict others' rights

Please also check the licenses of the individual source datasets.

Citation Information

If you use this dataset, please cite:

@dataset{tibetan_audio_merged_2024,
  title={Tibetan Audio-Sentence Merged Dataset},
  author={Your Name},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/your-username/tibetan-audio-sentence-merged}
}

Please also cite the original source datasets:

• lilgoose777/tibetan-speech-english-text-dataset
• lilgoose777/merged-tibetan-titung-goose
• Titung/tibetan-to-english-audio-dataset

Contributions

Contributions to improve this dataset are welcome! Please open an issue or pull request on the dataset repository.

Contact

For questions or issues regarding this dataset, please:

• Open an issue on the Hugging Face dataset page
• Contact: [your-email@example.com]

Acknowledgments

Special thanks to:

• Original dataset creators: lilgoose777 and Titung
• The Tibetan language community
• Hugging Face for hosting infrastructure

---

Version: 1.0
Last Updated: 2024
Status: Active