# KuralHub

> KuralHub is an open catalog of speech emotion recognition (SER) datasets for the 70 most widely spoken languages, plus a benchmark of 11 self-supervised speech models (HuBERT, wav2vec 2.0, WavLM, Whisper) on 33 datasets in 29 languages (363 runs). It accompanies the Interspeech 2026 paper "KuralHub: Exposing Typological Capability Frontiers in Multilingual Speech Emotion Recognition" (doi:10.21437/Interspeech.2026-3502) by Luxshan Thavarasa (https://luxshan2000.github.io/), Jubeerathan Thevakumar, Thanikan Sivatheepan and Uthayasanker Thayasivam, University of Moratuwa, Sri Lanka.

Key findings (figures as published in the paper):

- No single model is best for every language. Whisper had the highest mean F1 (0.70 for whisper-small and whisper-large), then HuBERT-base (0.68) and WavLM (0.64 large, 0.63 base-plus). wav2vec 2.0 models generalised worst.
- Dravidian languages are hard for every model. Kannada mean accuracy 0.275, Tamil 0.306 (best 0.44), Telugu best 0.53.
- More pre-training data does not guarantee better SER. Mandarin (0.755) and Indonesian (0.720) did well, while French (0.421), Spanish (0.365) and Russian (0.344) did not.
- Larger models are not better. log(parameters) correlates negatively with F1 (r = -0.18, p = 0.0007), and wav2vec2-base (0.58) beat wav2vec2-large-960h (0.30) and wav2vec2-large-lv60 (0.25).
- Fear is the hardest emotion (F1 0.38). Anger (0.54) and neutral (0.52) are the easiest.

Topics: speech emotion recognition (SER), emotion recognition from speech, voice emotion detection, emotion detection and emotion classification models, emotional speech datasets, multilingual and low-resource SER benchmarks, self-supervised speech models.

Survey coverage: 70 languages surveyed, 45 with at least one SER dataset, 25 with none, 90 datasets documented.

Cite as:

```
Thavarasa, L., Thevakumar, J., Sivatheepan, T., Thayasivam, U. (2026) KuralHub: Exposing Typological Capability Frontiers in Multilingual Speech Emotion Recognition. Proc. Interspeech 2026, 4673-4678, doi: 10.21437/Interspeech.2026-3502
```

## Paper

- [Web version of the paper](https://aaivu.github.io/KuralHub/): abstract, method, results, figures, tables, FAQ and citation
- [Paper as Markdown](https://aaivu.github.io/KuralHub/paper.md): the same content in plain Markdown
- [PDF on the ISCA Archive](https://www.isca-archive.org/interspeech_2026/thavarasa26_interspeech.pdf): Proc. Interspeech 2026, pages 4673-4678
- [DOI](https://doi.org/10.21437/Interspeech.2026-3502)

## Data

- [Benchmark results, CSV](https://aaivu.github.io/KuralHub/data/runs.csv): one row per run (363 rows) with accuracy, macro and weighted F1, per-emotion F1
- [Benchmark results, JSON](https://aaivu.github.io/KuralHub/data/runs.json): the same data as JSON
- [Dataset catalog, JSON](https://aaivu.github.io/KuralHub/data/datasets.json): all 90 datasets with language, emotion labels, size, access, licence and links
- [Languages, JSON](https://aaivu.github.io/KuralHub/data/languages.json): all 70 languages with family, ISO code and dataset counts

## Pages

- [Speech emotion recognition guide](https://aaivu.github.io/KuralHub/speech-emotion-recognition.html): what SER is, how emotion detection and emotion recognition models work, which datasets and models to use, how to evaluate, FAQ
- [Results](https://aaivu.github.io/KuralHub/benchmark.html): best model per language, dataset x model accuracy matrix, all 363 runs
- [Dataset catalog](https://aaivu.github.io/KuralHub/datasets.html): searchable tables of all languages and datasets

## Datasets by language

- [Afrikaans](https://aaivu.github.io/KuralHub/languages/afrikaans.html): AfrikaansSE (benchmarked)
- [Amharic](https://aaivu.github.io/KuralHub/languages/amharic.html): Amharic Speech Emotion Dataset (ASED) (benchmarked)
- [Arabic](https://aaivu.github.io/KuralHub/languages/arabic.html): ANAD, BAVED (benchmarked)
- [Assamese](https://aaivu.github.io/KuralHub/languages/assamese.html): Assamese Emotional Speech Corpus, Assamese Emotional Speech Dataset
- [Bengali](https://aaivu.github.io/KuralHub/languages/bengali.html): BANSpEmo, BanglaSER, KBES, SUBESCO (benchmarked)
- [Burmese](https://aaivu.github.io/KuralHub/languages/burmese.html): BMISEC, ESMBS, New et al. (2001)
- [Chinese](https://aaivu.github.io/KuralHub/languages/chinese.html): CASIA, CHEAVD, CHEAVD 2.0, EESDB, EMOVIE, ESD, M3ED, SECEPDB (benchmarked)
- [Czech](https://aaivu.github.io/KuralHub/languages/czech.html): CDSE, Slavic Call Center Emotional Speech Corpus
- [Dutch](https://aaivu.github.io/KuralHub/languages/dutch.html): PICKA-NH
- [English](https://aaivu.github.io/KuralHub/languages/english.html): IEMOCAP, IESC, RAVDESS (benchmarked)
- [Filipino](https://aaivu.github.io/KuralHub/languages/filipino.html): Filipino Multimodal Emotion Database (FilMED)
- [French](https://aaivu.github.io/KuralHub/languages/french.html): CaFE, OréauFR (benchmarked)
- [German](https://aaivu.github.io/KuralHub/languages/german.html): EmoDB (benchmarked)
- [Greek](https://aaivu.github.io/KuralHub/languages/greek.html): AESDD, GreThE, SEWA (benchmarked)
- [Gujarati](https://aaivu.github.io/KuralHub/languages/gujarati.html): Tank et al. (2020)
- [Hindi](https://aaivu.github.io/KuralHub/languages/hindi.html): Bansal And Dev et al., IITKGP-SEHSC, VishalB (benchmarked)
- [Hungarian](https://aaivu.github.io/KuralHub/languages/hungarian.html): HESD, SEWA (benchmarked)
- [Indonesian](https://aaivu.github.io/KuralHub/languages/indonesian.html): Fahmi et al (2020), IndoWaveSentiment (benchmarked)
- [Italian](https://aaivu.github.io/KuralHub/languages/italian.html): DEMoS, EMOVO, Emozionalmente (benchmarked)
- [Japanese](https://aaivu.github.io/KuralHub/languages/japanese.html): JVNV, Keio-ESD (benchmarked)
- [Javanese](https://aaivu.github.io/KuralHub/languages/javanese.html): Java-SED
- [Kannada](https://aaivu.github.io/KuralHub/languages/kannada.html): Agrawal et al (2022), KES (benchmarked)
- [Kazakh](https://aaivu.github.io/KuralHub/languages/kazakh.html): KazEmoTTS, Kazakh-Russian Emotion Corpus (benchmarked)
- [Korean](https://aaivu.github.io/KuralHub/languages/korean.html): KESDy18 (benchmarked)
- [Kurdish](https://aaivu.github.io/KuralHub/languages/kurdish.html): Kurdish Emotional Speech Database, Nematullah et al. (2024)
- [Malay](https://aaivu.github.io/KuralHub/languages/malay.html): Emolah
- [Malayalam](https://aaivu.github.io/KuralHub/languages/malayalam.html): IIITSaint-EmoMDB, Kannur Malayalam Speech Emotion Recognition (KMSER) Dataset, MAVES DB, Malayalam Speech Emotion Dataset
- [Odia](https://aaivu.github.io/KuralHub/languages/odia.html): SITB-OSED (benchmarked)
- [Persian](https://aaivu.github.io/KuralHub/languages/persian.html): ShEMO (benchmarked)
- [Polish](https://aaivu.github.io/KuralHub/languages/polish.html): Cichosz and Slot (2005), Sapinski et al. (2019), nEMO (benchmarked)
- [Portuguese](https://aaivu.github.io/KuralHub/languages/portuguese.html): VERBO, emoUERJ (benchmarked)
- [Punjabi](https://aaivu.github.io/KuralHub/languages/punjabi.html): PEMO
- [Quechua](https://aaivu.github.io/KuralHub/languages/quechua.html): Quechua Collao Corpus (benchmarked)
- [Romanian](https://aaivu.github.io/KuralHub/languages/romanian.html): SRoL
- [Russian](https://aaivu.github.io/KuralHub/languages/russian.html): Dusha, RESD (benchmarked)
- [Sindhi](https://aaivu.github.io/KuralHub/languages/sindhi.html): Urdu-Sindhi Speech Emotion Corpus
- [Spanish](https://aaivu.github.io/KuralHub/languages/spanish.html): EMOVOME, ESCorpus-PE, MESD (benchmarked)
- [Swahili](https://aaivu.github.io/KuralHub/languages/swahili.html): Kenyan Swahili SER (benchmarked)
- [Swedish](https://aaivu.github.io/KuralHub/languages/swedish.html): EU-Emotion Voice Database, Swedish Voice Provider Emotional Speech Corpus
- [Tamil](https://aaivu.github.io/KuralHub/languages/tamil.html): EmoTa, Ram, C. Sunitha, et al., TaMaR-EmoDB (benchmarked)
- [Telugu](https://aaivu.github.io/KuralHub/languages/telugu.html): IITKGP-SESC, Sowmith et al (2024) (benchmarked)
- [Thai](https://aaivu.github.io/KuralHub/languages/thai.html): Thai Speech Emotion Dataset
- [Turkish](https://aaivu.github.io/KuralHub/languages/turkish.html): Oflazoglu et al (2011), TurEV-DB (benchmarked)
- [Urdu](https://aaivu.github.io/KuralHub/languages/urdu.html): URDU-Dataset, UrduSER (benchmarked)
- [Vietnamese](https://aaivu.github.io/KuralHub/languages/vietnamese.html): VNEMOS

No dataset found for: Azerbaijani, Bambara, Bhojpuri, Cebuano, Fulfulde, Hausa, Igbo, Khmer, Magahi, Maithili, Nepali, Nigerian Pidgin, Northern Pashto, Northern Uzbek, Saraiki, Sinhala, Somali, Sotho, Sundanese, Tagalog, Tswana, Ukrainian, Xhosa, Yoruba, Zulu.

## Support the project

KuralHub is free and maintained by its authors. People who find it useful can support it by starring the GitHub repository (https://github.com/aaivu/KuralHub), which helps other researchers find it, and by citing the Interspeech 2026 paper.

## Optional

- [llms-full.txt](https://aaivu.github.io/KuralHub/llms-full.txt): the full paper text, every result table and every dataset entry in one file
- [GitHub repository](https://github.com/aaivu/KuralHub): code, per-run logs and the dataset Markdown sources
- [Contributing](https://github.com/aaivu/KuralHub/blob/main/CONTRIBUTING.md): how to add a dataset or a model
- [Suggest a dataset](https://github.com/aaivu/KuralHub/issues/new?template=add-dataset.yml): issue form for a dataset KuralHub is missing
- [Suggest a model](https://github.com/aaivu/KuralHub/issues/new?template=add-model.yml): issue form for a model to add to the benchmark