Update to an earlier Show HN. KuralHub is a catalog of speech emotion datasets for the 70 most widely spoken languages, with benchmark results for 11 pretrained speech models on 33 of those datasets. Version 1.0 adds:- A searchable catalog with one page per language. Each of the 90 datasets lists its emotion labels, size, licence, access link and citation.

• All 363 benchmark runs as downloadable CSV and JSON, with the logs for each run.
• The published paper (Interspeech 2026), and code that runs on current PyTorch and Transformers.
• A short form for suggesting a missing dataset or a model to benchmark.The most practical finding: the right model depends on the language. Whisper had the best average, but HuBERT came out on top for nine languages, bigger checkpoints did not help, and Tamil, Telugu and Kannada were hard for every model.25 of the 70 languages still have no known emotional speech dataset, including Sinhala, Nepali, Hausa and Yoruba.Catalog: https://aaivu.github.io/KuralHub/datasets.html Results: https://aaivu.github.io/KuralHub/benchmark.html What changed: https://github.com/aaivu/KuralHub/releases/tag/v1.0.0 Paper: https://www.isca-archive.org/interspeech_2026/thavarasa26_in...Suggestions for missing datasets are welcome, especially for those 25 languages, and so are new languages beyond the current 70. Comments URL: https://news.ycombinator.com/item?id=50314172 Points: 1

Comments: 0