I could conceivably be more interested in AI if what was being marketed was not just a tool that... (a) ensures you don't have to read/write anymore! (b) ensures you don't have to learn a language! (c) will plagiarize art, science, and scholarship! (d) will steal all your data, blogs, writing, etc!

I hope it's not "those curmudgeony phoneticians at Buffalo."

Oh yes, voice quality is another one of those things that remains measureable in 101 ways (and 'qualities' are often ranges of thresholds that might or might not map onto our existing phonetic labels).

I'm familiar with the paper - my take has always been about the findings (that duration is the strongest cue across languages) rather than the ways to operationalize the research. Though, the conflation of word-final/initial lengthening with stress is an important issue you point out.

I'll add, all the incentives are about making big claims on "Language" though. It's newsworthy, shareable, and simplifying. Grant review panels will favor general claims as "big impact", so the impetus is always to over-sell your findings and under-report language-specificity in results.

Most of what we are doing is exploratory because there are just too few speech scientists or phoneticians and too many topics and language communities. So, I guess I just worry that this is not possible unless we admit we are often working on language-specific phenomena.

My personal take is that you should not draw conclusions about "language" from a study on one language unless you have a priori evidence that languages do not vary along the dimension you are evaluating. So, where does that leave us in forming strong hypotheses?

Prosodic contrasts are variable across languages - stress is not the same in English, Spanish, and Mixtec. Should we as speech scientists ever be in the position to form overly strong hypotheses on the basis of work in languages that we are not actually testing in our studies? I would say "no."

The article states: "For example, researchers seem to have not yet agreed on how to acoustically measure cross-linguistically common phenomena such as word stress (e.g., Gordon & Roettger, 2017)." Perhaps there is no single way to analyze stress because it just differs across languages?

Second, the article concludes that there is a real need for more careful connections to be drawn between theoretical constructs and quantitative expectations, but this relies on having a robust space for formulating specific hypotheses. I don't see that as available for a lot of phonetic science.