End-to-end text-to-speech (TTS) systems allow for the generation of high-quality computer-generated speech without relying on expert-created modules. This paper outlines initial efforts to develop a Serbian end-to-end TTS system using the Tacotron architecture. Listening tests revealed that while Tacotron can produce natural-sounding synthesis when properly trained, it is prone to overfitting and requires extensive data to avoid frequent hallucinations and accent errors. The use of a vocoder proved to be crucial in overall speech quality. Although the level of Tacotron training is less critical, it still demonstrates easy overfitting with relatively small databases. Correct accents and the absence of artifacts and hallucinations are extremely important for listeners, and any issues in these areas result in significantly lower ratings. Despite being less expressive, a controllable standard DNN-based TTS with a standard front end receives better grades because it never hallucinates and rarely makes linguistic mistakes. Integrating expert knowledge from existing pipelines can further improve synthesis quality, especially in data-constrained scenarios.