Cantonese YouTube Audio-Caption Pairs for ASR

Ming030890's dataset contains Cantonese audio-caption pairs sourced from YouTube videos with manually provided captions. It was built by re-transcribing audio with SenseVoice and filtering segments to create a collection supporting ASR development. The dataset includes segments where ASR output matches original captions and segments with homophone or English word differences.

Use Cases

• Fine-tune Cantonese ASR models using the filtered audio-caption pairs where transcription matches the original manual caption.
• Analyze ASR error patterns by comparing segments where differences are only homophones or English words.
• Train speech recognition systems on a dataset specifically built from YouTube's Cantonese caption sources.

Strengths

• Dataset is built from YouTube videos with manually provided Cantonese captions, ensuring a foundation of human-verified text.
• Segments are filtered to include cases where ASR output is identical to the original caption, indicating high-quality alignment.
• Includes segments with specific error types like homophones or English words, useful for targeted ASR error analysis.

Limitations

• The dataset size, row count, and specific file formats are unknown, limiting assessment of scale and usability.
• Sample data and column structure are unavailable, preventing verification of data format and specific features.
• Potential bias towards YouTube content and the specific ASR model (SenseVoice) used for re-transcription.

Provenance

Source

YouTube videos with manually provided Cantonese captions.

Collection Method

Audio re-transcribed using SenseVoice, with segments filtered based on comparison to original captions.

Freshness

Last updated on 2025-07-20.