A Data-Centric Empirical Study of Fidelity-Prosody Trade-offs in Data Curation for Zhuang TTS

Training text-to-speech (TTS) systems on project-collected real-world recordings collected under multiple recording conditions that exhibit both environmental noise and atypical prosody is challenging, especially for tonal languages where prosodic structure directly affects intelligibility. This study adopts a multi-objective perspective on data curation that balances spectral fidelity and prosodic integrity and instantiates this perspective on a 24 hour screened subset drawn from a real-world Zhuang corpus of approximately 100 hours. Seven preprocessing strategies were compared under a controlled setting in which the training roster and model architecture were held fixed. The main empirical finding is that disabling internal silence compression while retaining trimming, filtering, and normalization (no_int_sil) yielded a substantial perceptual gain: Mean Opinion Scores (MOS) for overall naturalness improved from 3.17 for a minimally processed baseline to 3.72 (p < 0.05), even though a heavily cleaned configuration (full_method) attained the lowest Fréchet Audio Distance (FAD) among all systems. This pattern indicates that, in this project-collected Wuming Zhuang setting and under a VITS backbone, aggressive internal silence compression can act as a prosody-distorting operation: it improves apparent spectral cleanliness but perturbs boundary-related timing and pitch behavior that listeners rely on for naturalness.