What is prosody in speech?

Identical words mean different things depending on it. The stress moving one word along turns a statement into an accusation.

Questions are the obvious case in English. Rising pitch at the end signals one, and a system that flattens it turns every question into a command.

Synthesis has improved most here, and it is still where errors sound worst. A wrong emphasis makes a sentence feel insincere rather than merely odd.

Long sentences suffer most. The shape of a whole sentence needs planning, and a system generating as it goes has less to plan with.

Say your assistant confirms bookings with this line: "Your appointment, which is with Dr. Lee, has been scheduled for Thursday the 14th at 3:15 in the afternoon at our Elm Street office, so please arrive ten minutes early." You've handed the engine 30 words and one breath. It has to pick where the stress falls, and with that many clauses it often lands on "which" or "so" and skips the time.

Now split it into four lines: "You're booked with Dr. Lee. That's Thursday the 14th, at 3:15. We're at the Elm Street office. Can you come ten minutes early?" Each sentence you wrote carries one fact, so the stress has only one place to go. Your last line is a real question, and the pitch rises where a caller expects it to. Nothing about the voice changed between those two versions.

Call your own line and listen for three moments: a question, a number and an apology. Any question you hear should rise at the end. Numbers should come in groups with small gaps, the way you'd read a phone number to a friend. An apology read at the pace of a weather report sounds worse than no apology. If yours sounds cold, shorten it to "Sorry about that" and move on.

After that, open a week of your transcripts and search for callers saying "sorry?" or "can you repeat that?" Tally which assistant line came just before each one. If the same line shows up four or five times, you've found a rhythm problem, and rewriting that single sentence usually clears it.

Prosody matters less on short factual replies, and "We open at 8" is hard to get wrong. Long answers and anything with feeling in it are where your editing time pays back.

Short sentences are the practical lever, and they are yours to pull, because you write the script.

Where a specific line reads badly, rephrasing it usually works better than looking for a different voice.

The transcript shows where callers hesitated or asked for a repeat, which is often a prosody problem rather than a comprehension one.