this is an excellent parallel! my kid tends to homogeneously long sentences when discussing his latest pokemon cards and, boy, that's not super proficient communicative skills (tho' impressive use of subjunctive and conditional moods in those sentences!)

short utterances are as important as long utterances, so mean itself is not the right measure. The distribution of LU, however, is not very amenable to standard deviation (skewed, discrete). I remember trying both IQR and variance in a poisson model and both were better predictors than sd.

awesome thanks!

Those special effects always get me

Cool! Is there an open access version?

it didn't get funded :-/

yes, definitely. But I need to change the analysis in a registered report I inherited (long story), so I'm gearing up for the fight and more precise references will probably be needed :-)

when working on a second Puzzle of Danish grant (unfunded) we were playing with the idea that Danish might make things either easier for hearing loss (you developed good noise robust strategies), or harder (your system is already stretched). So yeah probably a lot to discover :-)

Super cool work! Do you know how that’d interact w hearing loss? Aging neurotypicals report similar problems, so I wonder about aging autistic individuals

I don’t know if it’s just me but I get called nicknames by my students in dk, which was quite upsetting at the beginning… perhaps because of the unscandinavian sounds in my name 🤷now I just keep aa amused gsheets . So far 27 and counting

great, I implemented it in my first script attempt, so this is a nice reassurance, thanks. I suspect a full model would need to use markov models to include sequential information within and between person (often several automated segments do all belong to one manual segment).

I will check them tonight, thanks! anything more specific on how to align automatically identified segments (sometimes with a generic SPEECH label) with manually annotated segments (which brings the many2one and one2many matching issue)?

I am now ready to face the more complex state-of-the-art algos. In particular, I'm trying to build a measurement error model, so assessing the match between automated and manual segments is quite crucial (ie which automated segments correspond to each manual segment). @dresen.bsky.social ?