ignificant progress has been made in developing text-to-speech (TTS) models. However, building TTS models specifically for children remains a challenge. The lack of child speech datasets and the difficulty in constructing such datasets have limited research in this area. Children often speak less clearly and exhibit significant variations in volume, rhythm, and the range of emotions they express. This study explores the utilization of a universal neural vocoder based on BIGVGAN and AutoVocoder for child speech synthesis. We trained the AutoVocoder on the adult LJ speech 1.1 dataset and then examined the effectiveness of using neural vocoders as a foundation for synthesizing conversational multi-speaker child speech. Our objective was to investigate how these vocoders can be fine-tuned and adapted to capture the distinct characteristics of child speech while minimizing the reliance on extensive child speech datasets. The experimental results demonstrated that the BIGVGAN vocoder outperformed others in synthesizing clear, natural-sounding conversational multi-speaker child speech. Despite challenges posed by the English MyST and Hungarian datasets used in this study, which included non-phonetic noise, indiscernible speech, and audio files of varying lengths, the AutoVocoder significantly enhanced the quality and clarity of the synthesized child speech. Preliminary findings indicate that the BIGVGAN model successfully generated high-quality synthesized child voices.