KazParaD: A Kazakh Paralinguistic Speech Dataset

KazParaD is a multi-source Kazakh speech corpus for prompt-controllable text-to-speech. It pairs speech with controlled-vocabulary style attributes (gender, age, pitch, emotion) and with 18 categories of non-verbal vocalizations (NVV). It was built to train and evaluate attribute-controllable TTS and NVV synthesis for a low-resource language.

The corpus has two subsets:

Subset	Utterances	Hours	Mean duration
demographic (paralinguistic)	69,817	62.31	3.21 s
nvv (nvv-18tags)	12,720	21.89	6.20 s
Total	82,537	84.20

The demographic subset

Source recordings are in-the-wild YouTube sourced Kazakh speech recordings. A subset of the labels is validated by native speakers.

The NVV subset

The NVV subset is synthetic: Kazakh sentences, each containing exactly one inline NVV tag (e.g. [laugh], [sigh], [cough]), were rendered by a text-to-speech system ElevenLabs across a small set of fixed speaker profiles spanning two genders and two age groups.

Audio format

• Container: WAV (PCM 16-bit), mono
• Sample rate: uniform 24 kHz

Attribute distributions

Gender (demographic subset)

Value	Count	% of labeled
M	31,892	45.7%
F	37,916	54.3%

Age category (demographic subset)

Value	Count	%
child	20,394	29.2%
young	17,620	25.2%
adult	19,188	27.5%
senior	12,615	18.1%

Pitch (demographic subset)

Value	Count	%
very high	28,129	40.3%
high	16,421	23.5%
medium	11,185	16.0%
low	7,722	11.1%
very low	6,360	9.1%

Emotion (demographic subset)

Value	Count	% of labeled
neutral	15,049	21.6%
sad	10,593	15.2%
happy	10,473	15.0%
angry	9,687	13.9%
fearful	9,410	13.5%
disgusted	8,346	12.0%
surprised	6,248	9.0%

NVV tags (18 categories)

Tag	Count	%
[long pause]	1,077	8.5%
[gasp]	1,047	8.2%
[sigh]	1,027	8.1%
[sniffle]	996	7.8%
[chuckle]	993	7.8%
[laugh]	945	7.4%
[groan]	895	7.1%
[cough]	889	7.0%
[whispers]	882	6.9%
[yawn]	796	6.3%
[singing]	617	4.9%
[stutters]	518	4.1%
[burps]	397	3.1%
[hiccups]	363	2.9%
[chewing]	319	2.5%
[smacks lips]	317	2.5%
[swallows]	316	2.5%
[tsking]	297	2.3%

Counts are over the 12,720 utterances of the NVV subset.

Licensing and terms

The annotations, transcripts, metadata and segment index are our own work and are released under CC BY-NC 4.0.

Citation

Paper is under review. A citation will be added here once it is published. Until then, please link to this repository.