Measuring phonetic correlates of stress (intensity in particular)

Question

It's normal enough for people to say that we can attribute the percept of 'stress accent' to certain acoustic correlates – usually higher relative fundamental frequency, intensity, and duration of syllables. It's easy enough to measure duration and compare that across 'stressed' and 'unstressed' syllables. But what about intensity?

There seem to be two things about intensity that I can't wrap my head around.

1. If intensity is measured in RMS, that means it's an average over time. So an intensity trace on Praat, for example, is working out some kind of average as it goes along from left to right. So what exactly should you measure then - the average RMS for the whole vowel? the highest RMS value in the intensity trace? is that intensity peak already somewhat flattened out if it's measured in RMS?

2. Once you have a measure, how do you compare them across syllables? Do you need to do a log transform of the RMS dB in order to make it a linear scale when you do a comparison, or would that not make sense because your ears hear logarithmically anyway?

I should point out that my research relates to Australian languages, not English, so the cues are considerably more subtle, and vowel quality does not change much between stressed and unstressed syllables.

Answer 1

I assume your goal is to describe the acoustics of "stress accent" in a language, distinguishing between stressed and unstressed syllables. Then you would treat amplitude and F0 in basically the same way. Both are window-defined functions: the (RMS) amplitude is a value in a certain window, as in F0. The difference is that RMS directly says "average", but still RMS amp. and F0 are reductions of waveform values in a window to a single number. The exact details of pitch and intensity windowings in Praat are not anything I know about, but you can select a chunk and get the same number of pitch and intensity values for the selection. With a sequence of numbers for a portion of speech corresponding to some segment or larger unit, you have to decide how to reduce those many points to few points.

The basic argument against picking highest or lowest points is that this is susceptible to artifacts, when what you want is reliable trends. One relatively minimalist method that I favor when dealing with vowel-centered measures is to divide the segment into thirds, and take the mean pitch / amplitude of that segment. That reduces a dozen points to 3 (and preserves some syllable-internal dynamics). Another is to compute a single value for the syllable: either the entire syllable (hoping that consonantal edge effects are minimal), or 50% of the syllable, in the center.

I don't see the merit of further massaging the computations with rescaling. The basic question should be "does this even matter?", so one of the fundamental issues that you will be concerned with is whether amplitude is even a significant factor, given F0 and formant values (that is, is amplitude an independent variable, or is it just an automatic consequence of something else). The reasoning here is that you want to know if there is something significant about amplitude as a correlate of stress in the physical signal. You can always discount a finding of significance if there is an overriding perceptual reason to think that the difference can't be noticed.