Stress position changes the meaning of a sentence
Aliases: contrastive focus · lexical stress · nuclear stress placement
What it is
The same segments, with stress on a different word, change what the sentence means. In-car TTS: “at the next light turn right” versus “at the next light turn right” — one prepares a turn, the other waits for that light. In English, lexical stress can also split noun from verb (REcord / reCORD). Contrastive focus marks what is new relative to the last turn. TTS often parks nuclear stress at the end by default, which answers a question the listener did not ask.
Why it happens
Listeners treat the prominent word as the answer to the current question under discussion. Stress on “right” and the question is which way; stress on “next” and the question is which light. Lexical stress regroups the rhythm of the segments; in English that is enough to change part of speech. Chinese leans on tone for word identity, but focus still uses intensity, duration, and pitch range to split “it is tomorrow we arrive” from “it is tomorrow we arrive.” Default end-weight is harmless in a neutral telling. Once the last turn already filled the final slot and the news sits mid-sentence, nuclear stress on the tail makes the old slot sound like a correction. The fault is not timbre. The voice chose the topic for the user.
Studying it
Gating: lengthen the same word or sentence in time slices and ask which alternative the contrast is (right versus left, this light versus the next). Or keep the segments and change only the F0 contour, then forced choice. Independent variable: which constituent carries the pitch accent. Dependent measure: meaning choice, not MOS.
Materials need a real minimal pair: two legal readings that fork only on prominence. “Sounds natural” with default final stress is not a comprehension pass. If listeners also see the text in capitals, the measure is reading.
Where it stops holding
When cabin noise masks stress cues, people fall back on word order; then write the contrast in words (“not this one, the next”) rather than relying on prominence. Fixed-stress languages use other focus markers. Highly conventional short commands have little focus structure to move. Correction of the last turn (“Tuesday, not Thursday”) is the home ground; a brand-new neutral informing often survives default final stress.
Applying it
- Mark information that is new relative to the last turn with emphasis on that word; do not let the engine dump nuclear stress at the end by default.
- On a correction, put stress on the changed slot: “Tuesday, not Thursday.”
- For navigation contrasts (left / right / next light), listen once without the script and confirm prominence lands on the contrast.
- How to check: one text, two emphasis patterns, two listener groups, different comprehension questions. If both groups answer the same, emphasis did not land.