Ask a learner what stress is and you get “say it louder”. That is about a third of it. Length and pitch move too — and in this corpus loudness is the most consistent of the three, while pitch barely moves at all.
English builds rhythm by making some syllables bigger than others. The word bigger hides three separate events: the syllable lasts longer, it is louder, and its pitch is usually higher. They come from different mechanisms — duration from timing, loudness from breath pressure, pitch from vocal fold tension — so they are not obliged to agree, and often they do not.
The numbers below say which of the three is reliable. Loudness moved the right way in 83% of the words, by +4.3 dB on median: roughly a doubling of what you hear as loudness. Duration moved in 64% of them, but by a large factor — the stressed vowel lasts 1.26× the median unstressed vowel of the same word. Pitch moved in 70% of them, by +0.88 semitone: audible, but not the thing carrying the signal.
That ordering has a practical consequence. If you listen for a pitch jump to tell you which syllable was stressed, you will miss a third of them.
Measured on 180 words
Six words where all three move together
The bars are the vowels of that word, at their measured durations; the filled one is stressed. Above each card are the other two parameters for that word. These six were picked because all three move the same way and by a wide margin — that is what makes them teachable, not what makes them typical.
Across 60 sentences, 180 words contained at least two vowels and exactly one stress mark, which is what makes the within-word comparison possible. 64% of those had a longer stressed vowel. The six above are the clearest, not the average.
Why the comparison stays inside one word
The obvious measurement — collect every stressed vowel in the corpus, compare it with every unstressed vowel — does not work, and we ran it first to find out. It returned 50 ms for both groups. The reason is that stress marks come from a pronunciation dictionary, and a dictionary marks the vowel of is as stressed exactly as it marks the vowel of demand. In running speech is is reduced; demand is not. The two groups were not stressed and unstressed vowels — they were content words and function words.
Comparing a stressed vowel with the unstressed vowels of its own word removes that. Both sit in the same word, from the same speaker, at the same tempo. What is left is the difference the stress made.
What it does not remove is vowel identity. An unstressed syllable in English usually reduces to a schwa, which is a short vowel by nature, so part of the 1.26× is stress and part is simply that two different vowels are being compared. That is unavoidable in real speech — and it is also, in a sense, the phenomenon itself: English stress is partly realised by not reducing.
The limits
Durations come from the forced aligner and are quantised to 20 ms frames, so a 50 ms vowel is two and a half frames of evidence. Loudness is the median intensity inside the vowel, 10 ms window. Pitch is in semitones above the median pitch of the same sentence, so a low voice and a high voice can be put on one axis.
And this is read speech. A narrator stressing a word for a listener makes it clearly bigger than someone stressing a word in conversation. The ordering of the three parameters should survive that difference; the sizes may not.