What is prosody?
Prosody is the rhythm and intonation of speech, covering where stress falls as well as how long a syllable is held. It carries information the words alone do not, such as whether a question has been asked or whether a figure matters. A synthesised line with correct words and wrong prosody is heard as wrong, which is why prosody is judged separately from clarity.
Where prosody carries the meaning the words leave out
A number read with the wrong phrasing is the most expensive prosody failure in financial calling. An account balance delivered as one undifferentiated string forces the caller to ask for it again, and an instalment date read without the pause a date needs is simply misheard. Prosody also governs whether an agent sounds rushed at the point where a caller is asked to agree to something. None of this appears in a transcript, so a deployment measured only on word accuracy will report that nothing is wrong.
- Prosody separates a question from a statement without changing a single word.
- Numbers and dates depend on phrasing to be heard correctly at all.
- A rushed delivery at a consent moment undermines the consent being recorded.
- Transcript based evaluation is blind to every prosodic failure.
How prosody works
- Text is normalised so that figures and dates are expanded into spoken form.
- Phrase boundaries are marked wherever a listener needs a pause.
- Emphasis is assigned to the tokens carrying the point of the utterance.
- The voice model generates duration and pitch contours over the phoneme sequence.
- Generated audio is checked against the intended phrasing for long numbers.
- Recurring problems are corrected in normalisation rather than in the model.
How prosody is evaluated
Prosody is assessed by listening, with the tests chosen for the utterances that matter commercially. Examine the following on generated audio:
- listener accuracy when a long number is played once
- mean opinion score for naturalness against a human reading
- correctness of question intonation on interrogative lines
- placement of phrase boundaries in dates and currency amounts
- perceived speaking rate at consent and disclosure moments
Limitations, risks and common failure modes
- Normalisation expands a figure correctly and places the pauses in the wrong positions.
- A voice tuned on English prosody applies English stress patterns to Hindi.
- Evaluation on isolated sentences misses how the agent sounds across a whole call.
- Latency work shortens pauses, which makes the agent sound impatient.