Taiwan Tongues: 10K-100K Hokkien ASR Audio Samples in WebDataset Format

Giving access to between 10,000 and 100,000 audio recordings and transcriptions for Hokkien speech recognition, published by adi-gov-tw in late 2024. It is organized into training and test subsets using the WebDataset format to facilitate high-throughput training in PyTorch and Hugging Face environments.

Use Cases

• Training automatic speech recognition models using audio and transcription pairs
• Benchmarking Hokkien STT performance using the test.tsv split
• Fine-tuning transformer-based audio models for Taiwanese dialects

Strengths

• 10,000 to 100,000 audio records
• WebDataset format for optimized I/O during deep learning training
• Pre-split into training and test sets for standardized benchmarking

Limitations

• Hokkien dialectal variation may lead to performance gaps across different regions of Taiwan
• License is listed as 'other', requiring manual verification of usage rights before commercial application

Provenance

Source

adi-gov-tw

Freshness

Last updated December 2025.

Geography

Taiwan