Prosodic error hurts comprehension more than a dull voice

What it is

Understanding a synthetic sentence cares more that prosody is right than that the voice is attractive. A sports watch says “in two hundred metres turn left” with nuclear stress on “metres” rather than “left,” and the wearer waits on distance instead of preparing the turn. The same segments with a drier timbre but intact phrase boundaries and focus often still parse; a lush timbre with the pitch accent on the wrong constituent sends attachment the wrong way. The comparison is two quality axes: plain but structurally right, versus pretty but structurally wrong. It is not the claim that stress by itself can flip a sentence into another sentence — that sits later, under rate and prosody.

Why it happens

Lexical identity rides on spectral cues and on expected prominence. Pitch accent and duration tell the listener which constituent is in focus, which way a modifier attaches, where a word boundary sits. Neural timbre can imitate a vocal tract closely and still dump nuclear stress on the end by default; the listener treats the tail as the news. A monotone that cuts in the right places still delivers word boundaries and default focus. Wrong prosody is a structural error in the message. MOS prefers the opposite: lush timbre scores; a misplaced focus need not be punished as unnatural, because it still “sounds like a person talking.” Comprehension items catch it.

Studying it

Degrade the same TTS two ways and compare comprehension, not MOS. One: vocode or flatten the spectrum, keep original F0 and duration. The other: keep original timbre, scramble or shift pitch accents. Materials: attachment-ambiguity sentences and focus-sensitive ones (“turn left” versus “the next light”). The dependent measure is a comprehension choice — who did what, which step to take — not a pleasantness score.

Listeners often call spectral damage “ugly but clear” and wrong stress “nice but I missed who did what.” If the lab lets people look at the text, attachment is already solved on paper and prosody is not what you measured.

Where it stops holding

In tone languages F0 is lexical as well as post-lexical; wrong prosody can destroy the word, not only the focus. For a brand jingle or an earcon, timbre is the product and this comparison does not apply. Highly conventional shorts (“please hold”) survive bad prosody. Informational speech — navigation, dosing, a recommended option — is the home ground. When noise wipes stress cues, people fall back on word order; then you need a lexical contrast, not prominence alone.

Applying it

• Put focus-sensitive items in the voice bake-off, not only “does it sound human.”
• For payloads (product names, turns, doses), prefer a plainer voice with correct phrasing over a celebrity voice that mis-stresses them.
• Mark new information with emphasis; listen once without the script and confirm nuclear stress lands on the word that must be acted on.
• How to check: after hearing, “what do you do next / which option was recommended.” If people quote the wrong constituent, it is prosody, not level.

Related

• Same group: M3.01.1 Naturalness and intelligibility are two different measures · M3.01.2 In noise, intelligibility comes first · M3.01.3 Too much naturalness raises expectations of understanding · M3.01.4 Proper names and digit strings fail first on intelligibility · M3.01.6 Long listening accumulates auditory fatigue
• Nearby: M3.02 Speech rate and prosody · M3.07 Rate, pauses and prosody
• Search terms: prosodic error costs more than dull timbre · misplaced prominence · nuclear stress