To address the unique linguistic characteristics of Cantonese in speech recognition, we propose WSYue-ASR-eval, a benchmark specifically designed for evaluating Cantonese ASR systems. It is tailored to assess model performance across diverse lengths, domains, and linguistic phenomena of Cantonese speech.
The test set annotations are provided by Beijing AISHELL Technology Co., Ltd.
Key features:
Annotated through multiple rounds of manual labeling
Includes rich tags such as text transcription, emotion, age, and gender
Covers Cantonese-English code-switching and multi-domain conditions
Enables comprehensive evaluation across varying speech lengths
WSYue-ASR-eval Subsets
Set
Duration
Speakers
Hours
Short
0–10 s
2861
9.46
Long
10–30 s
838
1.97
Total: 11.4 hours, with diverse speakers and scenarios.
