In a cup of tea, of is not /ɒv/. It is short, and its vowel has moved to the middle of the mouth. Only the first half of that is measurable at this resolution.
English keeps its grammar in a small closed set of words: articles, prepositions, pronouns, auxiliaries, conjunctions. On their own they carry almost no information, which is exactly why a speaker is free to compress them. A content word cannot shrink without becoming a different word — cup has to stay cup.
That asymmetry is what the measurement below picks up. We take every word in a sentence, divide its duration by the median duration of the words in the same sentence, and report the ratio. Content words cluster around 1.2–1.8. Function words that have been compressed fall to 0.3–0.5 — often 60–80 ms, which is three or four analysis frames.
The ratio is normalised inside the sentence on purpose. A slow speaker drags every word out; dividing by the sentence's own median removes the speaker's tempo and leaves only the relative compression, which is the thing that matters.
Duration, measured inside its own sentence
60 sentences · 981 measured words · 455 of them function words. Of those, 169 occurrences measured a ratio below 0.5 — that is 23 distinct words. Below are 16, lowest ratio first.
The shaded part of the envelope is the word's interval as the aligner timed it. The bars below compare this word's duration with the median word in the same sentence — that is the whole of what the ratio says.
The phonemes shown are not a transcription
Each example lists the phoneme sequence our grapheme-to-phoneme model produced for that word. It is what the aligner was told the word contains — not something recovered from this recording. Do not read it as “this is how it was said”.
The model is also inconsistent about reduction. For some words it returns a reduced form already (for → f ɚ, an → ə n, was → w ə z); for others it returns the full citation form with its stress mark (he → h ˈi, his → h ˈɪ z). Both kinds appear in the list below. That inconsistency is a property of the dictionary, not evidence about the speaker.
Why duration, and not vowel quality
The obvious way to show a weak form would be to measure the vowel: a schwa has a mid-central formant pattern, so find F1 and F2 and plot them. We tried it, and it does not survive 16 kHz conversational audio. On twelve vowels measured with Praat, five came out clearly wrong — tell returned F1 1031 / F2 1235 Hz where roughly 600 / 1800 Hz was expected. At 20 ms frames there is not enough pitch information for a reliable formant estimate.
Duration survives the same conditions because it does not depend on spectral detail. So that is what we report — rather than a formant chart that looks authoritative and is unreliable.