Reproduce “Training Text-to-Speech Model with Purely Synthetic Data: Feasibility, Sensitivity, and Generalization Capability” on OpenResearch

About this paper

Tingxiao Zhou, Leying Zhang, Zhengyang Chen, Yanmin Qian

This research systematically investigates the feasibility of training Text-to-Speech (TTS) models purely on synthetic data, demonstrating that models trained this way can achieve superior performance in speech quality and intelligibility compared to those trained on real data. The study quantifies the impact of various synthetic data attributes like speaker diversity, text richness, and noise levels on model performance and generalization capabilities.

Key insights

• Training TTS models on purely synthetic data can yield superior performance, particularly in intelligibility and overall speech quality, due to cleaner acoustic conditions compared to real data.
• Speaker diversity and text richness are critical synthetic data attributes, with increasing diversity consistently improving all performance metrics and text richness improving performance up to a point of sufficient phoneme coverage.
• TTS models are highly sensitive to noise and reverberation in training data; even small proportions of noisy speech or low SNRs (below 30 dB) significantly degrade performance.

Summary by alphaXiv.