WSYue-TTS-eval
收藏资源简介:
# WSYue-TTS-eval: Cantonese TTS Benchmark To address the unique linguistic characteristics of Cantonese in speech synthesis, we propose WSYue-TTS-eval, a benchmark specifically designed for zero-shot Cantonese TTS evaluation. It enables rigorous assessment of model generalization and robustness across diverse domains and challenging linguistic phenomena. ## Subsets - Base - 1,000 prompt–text pairs sampled from CommonVoice - Evaluates performance on real-world data distributions - Mainly daily conversational style - Coverage - Combines manually curated and LLM-generated texts - Spans multiple domains: daily life, news, entertainment, poetry - Covers diverse Cantonese linguistic phenomena: - Polyphonic characters - Tone sandhi - Code-switching - Proper nouns & numerals - Other challenging linguistic cases ### Highlights - Designed for rigorous evaluation of TTS systems - Tests generalization and robustness across diverse and complex Cantonese scenarios
# WSYue-TTS-eval:粤语文本转语音(Text-to-Speech, TTS)评测基准 为应对语音合成领域粤语独有的语言特征,我们提出WSYue-TTS-eval——一款专为零样本(Zero-shot)粤语文本转语音评测设计的基准数据集。该数据集可对模型在多元场景与复杂语言现象下的泛化能力与鲁棒性开展严谨评估。 ## 子集 - 基础集 - 从CommonVoice中采样的1000条提示-文本对 - 用于评估模型在真实世界数据分布下的表现 - 主体为日常会话风格 - 覆盖集 - 整合人工精选文本与大语言模型(LLM)生成的文本 - 覆盖日常生活、新闻、娱乐、诗歌等多个领域 - 涵盖多样化的粤语语言现象: - 多音字 - 连读变调 - 语码转换 - 专有名词与数字 - 其他高难度语言案例 ### 亮点 - 专为文本转语音系统的严谨评测而设计 - 可测试模型在多样且复杂的粤语场景下的泛化能力与鲁棒性




