GreekTTS-1.5

Abstract. Building upon our previous work in the Greek Text-to-Speech (TTS) space, this paper presents a significant leap forward in achieving state-of-the-art synthesis quality. The advance- ment of speech synthesis for less-resourced languages like Greek has been hampered by a lack of high-performance, accessible models. To address this definitively, we introduce a new state- of-the-art GreekTTS-1.5 system based on the more powerful Orpheus foundation model. Our approach utilizes Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning method, to effectively adapt this large-scale model to a custom, high-quality Greek speech corpus. The resulting system produces speech with exceptional naturalness and intelligibility, significantly outperforming existing baselines. This work offers a powerful, open-source GreekTTS-1.5 model and demonstrates an effective pathway for developing high-fidelity speech synthesis for other languages with limited resources.