Urdu speech emotion recognition datasets

KuralHub lists 2 speech emotion recognition datasets for Urdu, 2 of them openly available. Each entry below gives the emotion labels, size, recording setup, licence, where to get it and how to cite it. Urdu is also part of the benchmark. The best of 11 models reached 88.0% test accuracy (wav2vec2-base).

Language family: Indo-European (Indo-Aryan)
ISO code: ur
Spoken mainly in: Pakistan, and India (Uttar Pradesh)
Datasets listed: 2
In benchmark: Yes

Datasets at a glance

Dataset	Type	Samples	Speakers	Recording	Access	Licence
URDU-Dataset	Audio	400	38	Acted	Open
UrduSER	Audio	3500	10	Acted	Open	CC-BY 4.0

Benchmark results

Urdu is one of the 29 languages in the KuralHub benchmark. Each of the 11 frozen speech models was trained with a small classification head on the shared emotions and scored on a held-out test split. Over all 22 runs the mean test accuracy was 49.6% (median 45.5%). The best run reached 88.0% with wav2vec2-base. Compare with other languages.

**URDU-Dataset** (run ID `Urdu-Dataset`). Scores on the test split, best run highlighted.
Model	Family	Val acc.	Test acc.	Macro-F1	Weighted-F1	Logs
hubert-base-ls960	HuBERT	0.79	0.87	0.86	0.87	logs
hubert-large-ls960-ft	HuBERT	0.35	0.43	0.38	0.42	logs
wav2vec2-basebest	wav2vec 2.0	0.72	0.88	0.88	0.87	logs
wav2vec2-large-960h	wav2vec 2.0	0.26	0.34	0.25	0.30	logs
wav2vec2-large-lv60	wav2vec 2.0	0.18	0.35	0.27	0.30	logs
wav2vec2-xls-r-300m	wav2vec 2.0	0.29	0.41	0.36	0.38	logs
wav2vec2-xls-r-1b	wav2vec 2.0	0.23	0.35	0.26	0.32	logs
wavlm-base-plus	WavLM	0.72	0.76	0.75	0.76	logs
wavlm-large	WavLM	0.64	0.50	0.50	0.49	logs
whisper-small	Whisper	0.87	0.78	0.76	0.78	logs
whisper-large	Whisper	0.85	0.82	0.79	0.83	logs

**UrduSER** (run ID `UrduSER`). Scores on the test split, best run highlighted.
Model	Family	Val acc.	Test acc.	Macro-F1	Weighted-F1	Logs
hubert-base-ls960	HuBERT	0.53	0.51	0.50	0.51	logs
hubert-large-ls960-ft	HuBERT	0.31	0.32	0.27	0.26	logs
wav2vec2-base	wav2vec 2.0	0.44	0.46	0.44	0.45	logs
wav2vec2-large-960h	wav2vec 2.0	0.23	0.24	0.19	0.18	logs
wav2vec2-large-lv60	wav2vec 2.0	0.22	0.23	0.16	0.15	logs
wav2vec2-xls-r-300m	wav2vec 2.0	0.27	0.28	0.27	0.27	logs
wav2vec2-xls-r-1b	wav2vec 2.0	0.28	0.26	0.19	0.19	logs
wavlm-base-plus	WavLM	0.52	0.52	0.49	0.51	logs
wavlm-large	WavLM	0.44	0.45	0.44	0.44	logs
whisper-small	Whisper	0.59	0.56	0.55	0.56	logs
whisper-largebest	Whisper	0.60	0.60	0.59	0.60	logs

URDU-Dataset

Type: Audio
Emotion labels: Happy, Sad, Angry, Neutral
Size: samples - 400, speakers - 38 -
Speakers: 38
Recording: Acted
Access: Open
Licence: Not stated
File format: .wav
Website: Dataset Homepage
Paper: Paper Link
Download: Download, Download

The URDU dataset contains emotional utterances of Urdu speech gathered from Urdu talk shows. There are 400 utterances of four basic emotions in the book: Angry, Happy, Neutral, and Emotion. There are 38 speakers (27 male and 11 female). This data is created from YouTube. Speakers are randomly selected.

Used in the KuralHub benchmark (run ID Urdu-Dataset). See the results above.

Citation

@inproceedings{latif2018cross,
  title={Cross lingual speech emotion recognition: Urdu vs. western languages},
  author={Latif, Siddique and Qayyum, Adnan and Usman, Muhammad and Qadir, Junaid},
  booktitle={2018 International conference on frontiers of information technology (FIT)},
  pages={88--93},
  year={2018},
  organization={IEEE}
}

Source: datasets/Urdu/URDU-Dataset.md · Suggest an edit

UrduSER

Type: Audio
Emotion labels: Angry, Fear, Boredom, Disgust, Happy, Neutral, and Sad
Size: samples - 3500, speakers - 10
Speakers: 10
Recording: Acted
Access: Open
Licence: CC-BY 4.0
File format: .wav
Website: Dataset Homepage
Paper: Paper Link
Download: Download

The Urdu Speech Emotion Corpus (UrSEC) is a comprehensive dataset for Speech Emotion Recognition (SER) in Urdu, addressing the lack of large and diverse datasets in this language. It contains 3500 speech samples from 10 professional actors, representing seven emotional states: Angry, Fear, Boredom, Disgust, Happy, Neutral, and Sad. The samples are sourced from Pakistani Urdu drama serials and telefilms, ensuring natural and diverse speech. Each emotional class has 500 samples, with detailed metadata for easy organization and use in research. UrSEC is rigorously validated for accuracy and reliability, supporting advancements in Urdu SER.

Used in the KuralHub benchmark (run ID UrduSER). See the results above.

Citation

@misc{akhtar2024urdu,
  author    = {Muhammad Zaheer Akhtar and Rashid Jahangir and Qurat Ul Ain},
  title     = {UrduSER: A Dataset for Urdu Speech Emotion Recognition},
  year      = {2024},
  howpublished = {Mendeley Data, V3},
  doi       = {10.17632/jcpfjnk5c2.3}
}

Source: datasets/Urdu/UrduSER.md · Suggest an edit

Know another Urdu speech emotion dataset, or a newer version of one listed here? Suggest it or open a pull request.

This page is generated from datasets/Urdu/ in the KuralHub repository. If you use a dataset, cite its authors. If KuralHub helped you find it, please also cite the KuralHub paper.