As I've learned over time (and other people in these comments have clarified) it turns out that evaluating "quality" of Text To Speech is somewhat dependent on

As I've learned over time (and other people in these comments have clarified) it turns out that evaluating "quality" of Text To Speech is somewhat dependent on the domain in which the audio output is being used (obviously with overlaps), broadly: * accessibility * non-accessibility (e.g. voice interfaces; narration; voice over) The qualities of the generated speech which are favoured may differ significantly between the two domains, e.g. AIUI non-accessibility focused TTS often prioritises "realism" & "naturalness" while more accessibility focussed TTS often prioritizes clarity at high words-per-minute speech rates (which often sounds distinctly non-"realistic"). And, AIUI espeak-ng has historically been more focused on the accessibility domain.