Voice Identification with Artificial Corpus for Serbian language

Text-to-speech (TTS) technologies have advanced significantly in recent years, yet languages with limited resources, such as Serbian, still lack systems capable of generating diverse and controlled synthetic speech suitable for research and downstream applications. This thesis addresses this gap by developing a multi-speaker Serbian TTS system capable of producing synthetic voices that represent distinct demographic profiles defined by gender and age. The central research question of this work is whether a modern neural TTS architecture can learn and reproduce demographic speaker characteristics when trained solely on a synthetic dataset. To investigate this, a complete processing pipeline was designed , including dataset construction, preprocessing, model configuration, training, inference, and evaluation. Written text was paired with synthetic audio generated using commercialTTS services, forming a balanced corpus for six speaker groups. The Glow-TTSarchitecture, implemented within the Coqui TTS framework, was used as the core model, supported by trainable speaker embeddings and phoneme-based input representations. The system was evaluated using both objective measures (MOSNet, PESQ, STOI)and a comprehensive subjective listening test assessing intelligibility, naturalness, and demographic perception. The results show that the model produced intelligible synthetic speech with consistent quality across speakers and preserved gender cues reliably. Agerelated characteristics were captured only partially: young and adult profiles were recognisable, while elderly voices were frequently perceived as adult, reflecting the limitations of purely synthetic training data and reduced prosodic variation. Overall, the thesis demonstrates that controlled demographic voice synthesis for Serbian is feasible even in the absence of natural speech recordings. The designed pipeline lays the groundwork for future research in multilingual demographic modelling, enhanced prosodycontrol, and dataset generation for downstream tasks such as speaker classification and fairness evaluation in machine learning systems.