Read a sentence aloud and the pitch does not stay put. It starts near the top of your range and works downward, and the last third of a second decides whether the sentence sounds finished.
English statements have a shape. The first stressed syllable sets a high point, each following phrase steps down, and the final syllable lands near the bottom of the speaker's range. Questions and unfinished clauses break that shape — which is how a listener separates she left from she left? with no punctuation to help.
In this corpus 55 of 60 sentences end on a falling pitch, 92% of them. The median fall is -4.5 semitones per second across the final third of a second, and the start of a sentence sits 8.3 semitones above its end — most of an octave. The whole sentence moves inside a range of 7.5 semitones between its 10th and 90th percentile.
The 5 that rise are not errors. They are sentences cut out of the middle of a paragraph, where the clause has not finished. A rising ending is what an unfinished clause is supposed to sound like.
Measured on 60 sentences
Four sentences, four shapes
Each graph is that recording's pitch track: 10 ms frames, vertical axis in semitones above the sentence's own median, so the dashed line at zero is that speaker's middle. The shaded end is the window the slope is fitted over. Breaks in the line are unvoiced sounds — a consonant, a pause — and the line is broken there on purpose rather than drawn through.
its slope is closest to the corpus median — this is what an ordinary sentence looks like
the steepest fall in the corpus
almost no fall at all
one of the 5 that end higher than they started
How the line is measured
Pitch comes from Praat's autocorrelation method, 10 ms step, 75–500 Hz. That method occasionally locks onto twice or half the true period, which appears as a one-octave jump in an otherwise smooth track. We repair exactly those frames: a frame more than eight semitones from the median of its own ±150 ms neighbourhood is folded back by a power of two. Nothing else is touched. Before that step one sentence reported a pitch range of 18 semitones while the corpus median was 7.5 — the sentence was ordinary, the measurement was not.
The slope is a straight-line fit over the final max(350 ms, 25% of the sentence). Reporting it per second rather than in total keeps short and long sentences comparable. Everything is in semitones relative to the median pitch of the same sentence, which removes the speaker's voice from the picture.
What this cannot show
There are no questions in this corpus. LibriSpeech is volunteers reading novels and the transcripts carry no punctuation, so nothing tells us which sentence was meant as a question. Every shape here is a statement, or a clause inside one. The rising examples are the corpus's own accidents, not a designed contrast.
It also cannot tell you what the speaker meant. Pitch moves with emphasis, with hesitation and with the paragraph the sentence came from, and this recording is one sentence removed from all of that.