Hakka Speech Recognition Dataset for Taiwanese Languages

Hakka-language audio recordings and transcriptions form a pre-training dataset for the Taiwan-Tongues-ASR-CE project. The dataset is packaged in WebDataset format for direct use with PyTorch and Hugging Face libraries. It was created by the adi-gov-tw organization and last updated in December 2025.

Use Cases

• Train automatic speech recognition models on Hakka audio files and corresponding text transcriptions.
• Fine-tune pre-trained ASR models using the provided training and test splits (train.tsv, test.tsv).
• Benchmark model performance on the separate test set of Hakka speech samples.
• Develop multilingual ASR systems incorporating Hakka using the WebDataset-formatted tar archives.

Strengths

• Data is structured for machine learning pipelines with dedicated training and test subsets.
• Format compatibility with WebDataset, PyTorch, and Hugging Face Datasets simplifies loading.

Limitations

• Unknown sample size and audio duration limit statistical analysis.
• Lack of column details prevents assessment of feature richness or label consistency.

Provenance

Source

adi-gov-tw organization on Hugging Face.

Collection Method

Pre-training data collected for the Taiwan-Tongues-ASR-CE project, method unknown.

Time Range

null

Freshness

Last updated in December 2025.

Geography

Taiwan, focusing on the Hakka language.