Human Ears Over Loss Curves: The Real Driver of TTS Fine-Tuning

While numerical metrics are standard, they often fail to capture the perceptual nuances essential for high-quality speech synthesis. This exploration shows how iterative human feedback directly informs critical engineering decisions, such as learning rate scheduling and data selection, to master specific phonetic patterns.