Why WER, CER, MOS, and validation loss often fail to capture whether a text-to-speech model actually sounds better.