prosody is the pattern of pitch, timing, stress, loudness, and pauses in speech. it's why the same words can sound like a question, a correction, a warning, or reassurance. in text to speech (tts), prosody shapes how a line is delivered around the words.
how does prosody work?
speech researchers describe prosody through several acoustic features that work together.
pitch rises and falls form intonation, which helps signal how a phrase should be understood. the duration of sounds and the placement of pauses shape rhythm. stress makes a syllable or word stand out through a combination of pitch, duration, and loudness.
these patterns can operate across one syllable, a word, or a full sentence. a tts model may pronounce every word correctly and still sound awkward if it stresses the wrong detail or pauses in the middle of a phrase.
prosody example
take the line, "your appointment is on friday."
stress your, and the speaker may be distinguishing this appointment from someone else's.
stress friday, and the speaker may be correcting the date.
let the pitch rise at the end, and an english listener may hear a request for confirmation.
let it fall, and the line sounds more like a settled answer.
the words stay the same. the prosody changes the listener's reading of them.
this matters in a voice agent. if a caller asks, "did you say thursday?", the reply should make friday easy to hear. a flat delivery can bury the correction even when the transcription is accurate.
prosody vs voice design
the three ideas affect different parts of a tts result:
voice design describes the speaker's overall sound, including accent, age range, pitch, timbre, and speaking style.
prosody describes how a particular utterance moves through stress, pitch, rhythm, intensity, and pauses.
pronunciation describes how the sounds in a word are produced.
a voice can keep the same identity while its prosody changes from calm to urgent. it can also have convincing prosody and still mispronounce a person's name.
why does prosody matter for voice agents?
rumik silk offers two kinds of control over delivery.
mulberry voice model uses a voice description. a brief such as "an indian english voice, normal pitch, conversational pacing, calm delivery" gives the model broad direction for pitch, pace, and emotion across the speech.
muga uses tone tags. a tag such as [happy], [sad], or [neutral] changes how the whole line is spoken. in rumik's muga, [happy] gives the line a bright, smiling, mid-energy delivery. [excited] makes it louder and faster, with a higher pitch.
you can compare the same line in rumik's playground while changing one instruction at a time. keep the text and speaker fixed, then change the description, or tone tag. that makes the audible effect easier to identify.
use lines that force the tts model to make clear choices. a useful test set includes:
a greeting followed by a question
a correction such as, "not tuesday. your appointment is on friday"
a line containing a date, time, or amount
a longer reply with two or three sentence breaks
a code-switched line used in the real call flow
listen for the detail that receives stress, the pause before important information, and the pitch at the end of a question. check that longer replies do not become flat or rushed. in code-switched speech, listen for rhythm or emotion changing abruptly at the language boundary.
word error rate can show if a speech recognizer recovered the intended words. it does not tell you if the correction sounded firm, the question sounded open, or the pause landed in the right place. prosody still needs a listening test with the kinds of lines the voice agent will actually say.