<p>Over the long-time efforts, there are still some shortcomings in speech synthesis technology. The major drawbacks include unnatural pronunciation, stiff cross-lingual synthesization, synthesized noisy artifacts, and unsatisfactory synthesizing speed. Up to now, customizing a speech synthesis system to capably mimic the timbre, intonation, rhythm, and prosody of specified speakers remains challenging. To address the above issues, this study proposes an advanced text-to-speech model. The model integrates an autoregressive mechanism, the Bidirectional Encoder Representations from Transformers (BERT), the Chinese Hidden-Unit BERT (HuBERT), and self-supervised learning within the Variational Inference with adversarial learning for end-to-end Text-to-Speech (VITS). Combining these three core components, our proposed model is named ABS-VITS. The inference-adapted design improves the original VITS, while the fine-tuned BERT enhances timbre reconstruction. In addition, a randomized duration prediction mechanism generates speech with diverse rhythms, enabling more realistic and high-fidelity human-like synthesis. By effectively organizing the pre-trained models and fine-tuned modules, ABS-VITS can straightforwardly achieve high fitness in producing speech waveforms from given texts. Particularly, ABS-VITS is obviously superior to the vanilla (prototype) VITS in terms of energy consumption required for fine-tuning. Moreover, the framework of the proposed network allows users to freely choose their desired target speaker, thereby enhancing the naturalness of synthesized speech after rapid fine-tuning. Consequently, ABS-VITS can be an efficient tool for personalized prosodic cross-lingual speech synthesis with vivid emotional expression.</p>