KazEGA

KazEGA is a Kazakh-language speech corpus for paralinguistic classification. Each utterance is annotated for three speaker attributes — emotion, gender, and age group.

The corpus contains 96,582 utterances. The training set comprises 88,977, the validation set 3,978, and the test set 3,627 utterances.

Tasks

• Emotion recognition — seven classes: angry, disgusted, fearful, happy, neutral, sad, surprised.
• Gender classification — two classes: female, male.
• Age-group classification — four classes: child, young, adult, senior.

Emotion distribution

Emotion	Train	Validation	Test	Total
Angry	10,290	533	389	11,212
Disgusted	8,128	85	69	8,282
Fearful	10,078	232	143	10,453
Happy	10,532	574	485	11,591
Neutral	33,169	1,817	1,931	36,917
Sad	10,922	445	468	11,835
Surprised	5,858	292	142	6,292
Total	88,977	3,978	3,627	96,582

Gender distribution

Gender	Train	Validation	Test	Total
Female	36,715	1,617	1,353	39,685
Male	52,262	2,361	2,274	56,897
Total	88,977	3,978	3,627	96,582

Age-group distribution

Age group	Train	Validation	Test	Total
Child	14,996	790	342	16,128
Young	16,060	707	698	17,465
Adult	44,072	2,336	2,398	48,806
Senior	13,849	145	189	14,183
Total	88,977	3,978	3,627	96,582

Data fields

Each example provides the speech recording (16 kHz audio) together with its emotion, gender, and age-group labels, and, where available, an orthographic transcription of the utterance.

Data collection and processing

Source audio was collected from publicly available Kazakh-language YouTube content (news, interviews, podcasts, documentaries, educational videos, animations). Long-form audio was segmented with voice activity detection (VAD). Speaker turns were assigned with speaker diarization. Clips shorter than 1 s or longer than 20 s were discarded. Candidate utterances were pre-labeled automatically and then verified and corrected by native Kazakh speakers through a custom annotation platform. Training and validation splits include noise-augmented samples (35.1% and 12.8%). The Test Split is real and unaugmented.

Licensing and ethics

The source videos carry the Standard YouTube License, which does not grant redistribution rights for extracted audio. This dataset is therefore released for non-commercial research only under gated access with a research-use agreement. The CC-BY-NC-4.0 license applies to the annotations. No rights over the underlying audio are claimed.

Utterances carry no speaker identities, names, or transcript-level personal data. The corpus includes child speech, subject to the same access and takedown terms.

Takedown. Speakers or rights holders may request removal of their material by opening a discussion in this repository.