KazParaD: Paralinguistic and Non-Verbal Vocalization Corpus
for Kazakh Speech Synthesis

Prompt-based controllable text-to-speech (TTS) has recently become an effective approach for voice styling. Extending these methods to low-resource languages is limited by data scarcity, and the systematic integration of paralinguistic attributes and non-verbal vocalizations (NVVs) has not been explored. We introduce KazParaD, a Kazakh speech corpus of over eighty hours with controlled-vocabulary style annotations, combining in-the-wild recordings validated by native speakers with a synthetic subset covering 18 NVV categories. On this resource we build KazGenericTTS, a prompt-controllable system with a discrete, non-autoregressive architecture initialized from a large language model (LLM), whose unified token space treats NVV tags as decoding targets and thus avoids multi-stage style encoders. We evaluate it with subjective listening tests and objective probes validated against Kazakh human labels. The model outperforms zero-shot multilingual baselines in NVV realization and speaker gender control, and in-language adaptation reduces recognition error below that of the noisy ground-truth recordings while maintaining multi-attribute control.