Question about the CER of analysis-synthesis and TTS system

[Liujingxiu23]

In the paper, as shownd in the Figure 5 and claimed in part 6.3, I did not understand, why cer of LLM-based TTS system is more lower than the cers of audio from analysis-synthesis? No matter PQ, RQ or OPQ. It seems not in line with intuition

[hhguo]

This is because they are from different test sets. The analysis-synthesis test set presents higher diversity in noise level, recording environment, timbre, and speaking style, etc..., making it more challenging for perfect reconstruction. The TTS test set uses relatively clean audio as references.