Titung/tibetan-to-english-audio-dataset

Tibetan to English Audio Dataset

Dataset Description

A Tibetan speech recognition dataset with transcriptions and English translations containing 1,642 audio samples.

Dataset Summary

This dataset contains Tibetan speech recordings with:

• Tibetan transcriptions in native script
• English translations
• High-quality audio files in WAV format

Total Samples: 1,178
Total Size: ~1.1 GB
Audio Format: WAV

Languages

• Source Language: Tibetan (བོད་སྐད་)
• Target Language: English

Dataset Structure

Data Fields

• audio: Audio file containing Tibetan speech (automatically loaded as audio array)
• transcription: Tibetan transcription of the speech in native script
• sentence: English translation of the Tibetan transcription

Data Splits

This dataset contains a single train split with 1,178 examples.

Usage

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("Titung/tibetan-to-english-audio-dataset")

# Access the train split
train_data = dataset['train']

# View first example
print(train_data[0])

Example Usage with Audio

from datasets import load_dataset

# Load dataset
dataset = load_dataset("Titung/tibetan-to-english-audio-dataset")

# Get first example
example = dataset['train'][0]

# Audio is automatically loaded
print(f"Audio array shape: {example['audio']['array'].shape}")
print(f"Sample rate: {example['audio']['sampling_rate']} Hz")
print(f"Transcription: {example['transcription']}")
print(f"Translation: {example['sentence']}")

Fine-tuning Whisper for Tibetan ASR

from datasets import load_dataset
from transformers import WhisperForConditionalGeneration, WhisperProcessor

# Load dataset
dataset = load_dataset("Titung/tibetan-to-english-audio-dataset")

# Load Whisper model
processor = WhisperProcessor.from_pretrained("openai/whisper-small")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")

# Prepare data
def prepare_dataset(batch):
    audio = batch["audio"]
    batch["input_features"] = processor(
        audio["array"], 
        sampling_rate=audio["sampling_rate"],
        return_tensors="pt"
    ).input_features[0]
    batch["labels"] = processor.tokenizer(batch["transcription"]).input_ids
    return batch

dataset = dataset.map(prepare_dataset, remove_columns=dataset.column_names)

# Now ready for training!

Dataset Creation

Source Data

The dataset was created from Tibetan speech recordings with manual transcriptions and professional English translations.

Annotations

• Transcriptions: Native Tibetan transcriptions in traditional Tibetan script
• Translations: English translations of the Tibetan text

Audio Processing

• Format: WAV
• Quality: High-quality audio recordings
• Total size: ~1.1 GB

Use Cases

This dataset is suitable for:

• Tibetan automatic speech recognition (ASR)
• Tibetan-to-English speech translation
• Tibetan language processing research
• Low-resource language model training
• Multilingual speech recognition systems
• Fine-tuning Whisper models for Tibetan

Limitations

• Single speaker or limited speaker diversity (please verify based on your data)
• Specific domain or topic coverage (please specify based on your data)

Additional Information

Licensing Information

This dataset is released under cc-by-4.0 license.

Citation Information

@dataset{tamang_tibetan_english_2026,
  author       = {Anil Tamang},
  title        = {Tibetan to English Audio Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/Titung/tibetan-to-english-audio-dataset}
}

Contributions

This dataset was created to support Tibetan language processing and speech recognition research.

Contact

For questions or issues, please open an issue on the dataset repository.

---

Dataset Statistics:

• Total samples: 1,178
• Total size: ~1.1 GB
• Format: WAV audio with automatic loading
• Dataset Viewer: ✅ Enabled