Fine-Grained Prosody-Controllable Lao Speech Synthesis Guided by Natural Language
摘要
Natural language-driven Lao speech synthesis faces three major challenges: first, the scarcity of large-scale high-quality annotated corpora; second, the modeling difficulty caused by the complex syllabic structure and tonal system of the Lao language; and third, the insufficient granularity of existing methods in rhythm regulation. To address these issues, this paper first constructs a Lao dataset with automatically annotated natural language descriptions and proposes a high-fidelity controllable speech synthesis method. This method utilizes natural language descriptions for global style control and performs fine-grained local modeling through word-level prosody latent space, achieving unified control over the expressiveness of the synthesized speech. Experimental results show that the Lao speech synthesized by this method achieved a MOS score of 3.69, validating the effectiveness of collaborative rhythm regulation in enhancing the expressiveness of Lao speech.