Odia speech emotion recognition datasets

KuralHub lists 1 speech emotion recognition dataset for Odia. Each entry below gives the emotion labels, size, recording setup, licence, where to get it and how to cite it. Odia is also part of the benchmark. The best of 11 models reached 83.0% test accuracy (wavlm-large).

Language family: Indo-European (Indo-Aryan)
ISO code: or (ISO 639-1), ori (ISO 639-2)
Spoken mainly in: India (primarily in Odisha state)
Datasets listed: 1
In benchmark: Yes

Datasets at a glance

Dataset	Type	Samples	Speakers	Recording	Access	Licence
SITB-OSED	Audio	12,110	20	Studio	Open

Benchmark results

Odia 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 11 runs the mean test accuracy was 64.4% (median 64.0%). The best run reached 83.0% with wavlm-large. Compare with other languages.

**SITB-OSED** (run ID `SITB-OSED`). Scores on the test split, best run highlighted.
Model	Family	Val acc.	Test acc.	Macro-F1	Weighted-F1	Logs
hubert-base-ls960	HuBERT	0.76	0.75	0.74	0.75	logs
hubert-large-ls960-ft	HuBERT	0.61	0.62	0.62	0.62	logs
wav2vec2-base	wav2vec 2.0	0.65	0.64	0.60	0.62	logs
wav2vec2-large-960h	wav2vec 2.0	0.44	0.43	0.39	0.40	logs
wav2vec2-large-lv60	wav2vec 2.0	0.29	0.28	0.20	0.21	logs
wav2vec2-xls-r-300m	wav2vec 2.0	0.50	0.51	0.48	0.49	logs
wav2vec2-xls-r-1b	wav2vec 2.0	0.54	0.58	0.53	0.55	logs
wavlm-base-plus	WavLM	0.79	0.78	0.77	0.77	logs
wavlm-largebest	WavLM	0.82	0.83	0.83	0.83	logs
whisper-small	Whisper	0.83	0.83	0.83	0.83	logs
whisper-large	Whisper	0.84	0.83	0.83	0.83	logs

SITB-OSED

Odia Speech Emotion Database (SITB-OSED)

Type: Audio
Emotion labels: Anger, Surprise, Happy, Sadness, Disgust, Fear
Size: 12,110 utterances, 20 speakers (10 male, 10 female)
Speakers: 20
Recording: Studio
Access: Open
Licence: Not stated
File format: .wav (16-bit, 22.05 kHz, stereo)
Website: SPEAL Website
Paper: IEEE Xplore
Download: Download
Contact: mail

SITB-OSED is the first open-source large Odia speech emotion database developed by Silicon Institute of Technology, Bhubaneswar. The dataset contains recordings from 20 professional Odia native speakers covering five major Odia dialects (Cuttacki, Baleswari, Berhampuri, Sambalpuri, and Phulbani). Each utterance ranges from 3.5s to 8s and was recorded in a quiet environment. The dataset represents balanced coverage of six primary emotions across all major Odia dialects.

Used in the KuralHub benchmark (run ID SITB-OSED). See the results above.

Citation

@INPROCEEDINGS{10089254,
  author={Maji, Bubai and Swain, Monorama},
  booktitle={2022 IEEE Asia-Pacific Conference on Computer Science and Data Engineering (CSDE)},
  title={SITB-OSED: An Odia Speech Emotion Dataset},
  year={2022},
  volume={},
  number={},
  pages={1-5},
  keywords={Computer science;Emotion recognition;Speech analysis;Databases;Machine learning;Data engineering;Recording;speech dataset;speech emotion recognition;deep learning},
  doi={10.1109/CSDE56538.2022.10089254}}

Source: datasets/Odia/SITB-OSED.md · Suggest an edit

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

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