Glossary · what detection measures

Prosody

The music of speech rather than its words. It carries most of what a sentence means beyond its dictionary content, and it used to be where machine speech gave itself away.

Definition

Prosody is the layer of speech carried above individual sounds: timing, rhythm, loudness, stress placement and the rise and fall of pitch. It signals emphasis, mood, whether an utterance is a question, and when a speaker has finished. Detection uses it because natural prosody varies in ways models reproduce imperfectly.

It is where the meaning lives

Take a single sentence — I never said she took the money — and move the stress. Seven placements, seven different accusations, identical words. None of that difference is in the vowels and consonants; all of it is in prosody.

Linguists call these features supra-segmental, meaning they are carried across the individual sounds rather than inside them. They include the pitch contour of a phrase, where pauses fall and how long they are, whether tempo speeds up under excitement, and which syllables get extra length or loudness. They also do interactional work: signalling that you have finished a turn, or that you are about to add something, or that you doubt what you just heard.

Why it was a detection signal, and why it faded as one

Earlier synthesis produced prosody that was too regular. Pauses of uniform length, pitch contours that repeated their shape, stress landing by rule rather than by meaning, and the eventual result was speech that sounded flat or oddly emphatic. Human listeners noticed, and the advice of the period — listen for a monotone, listen for missing breaths — was reasonable at the time.

Current systems model this layer directly and can be conditioned on a reference performance, so the audible version of the cue has largely gone. That is the central point of our guide to listening: acting on cues that stopped working is worse than not listening at all, because it keeps you on the call while you deliberate.

What survives is machine-measurable rather than audible, and it needs material. Prosodic structure is a statistical property of a whole utterance, so a four-second clip may contain too few pauses, stressed syllables and phrase boundaries for any statement about its distribution to mean much. This is one of the reasons very short recordings return unclear rather than a confident answer.

A concrete case

An investigator submits a six-second voice note: a name, a figure, and an instruction. It is emphatic, clipped and delivered in one breath, and it contains perhaps two phrase boundaries in total.

There is nothing wrong with the recording, and there is very little prosody in it to analyse. A longer sample of the same speaker, on the same call, would carry far more. Where a submission is being chosen, more speech of ordinary quality beats a short excerpt of good quality, which is the opposite of most people’s instinct.

Commonly confused with: timbre, and with accent

Timbre is the quality that makes a voice recognisably one person’s — the resonant character of their vocal tract. It is a property of the frequency structure and belongs to spectral analysis. Prosody is about how that voice is deployed in time. Two people can share a prosodic style and sound nothing alike; one person can keep their timbre and change their prosody completely between a lecture and an argument.

Accent is a third thing again, and overlaps with both: it involves how sounds are articulated as well as characteristic rhythm and intonation patterns. Describing a synthetic voice as having “flat prosody” when what is meant is an unfamiliar accent is a common and unhelpful error, and one that lands unevenly on speakers whose accent is not the one a system was mostly trained on.

What Truthring does with it

Prosodic measurements are one family among several in the pipeline, and their contribution scales with how much usable speech the clip contains. They are also close to useless in one specific case worth naming: in voice conversion, a real person supplies the entire performance and only the vocal identity is replaced, so the prosody in the file is genuinely human and nothing about it should look wrong.

Being explicit about which features stop contributing under which conditions is what an honest confidence figure is made of. Where the signal is thin, the number comes down. Where too much is thin at once, the verdict is unclear, which we treat as a real answer rather than a failure.

FAQ

Questions this term raises

Is flat or robotic delivery still a sign of AI speech?

Rarely. Current systems model rhythm and intonation directly and can carry hesitation and emotional colour convincingly. On a phone line the codec flattens genuine speech too, so a person under stress often sounds more mechanical than a generated clip does.

Is prosody the same as tone of voice?

Close, in ordinary usage. Prosody is the technical description of what produces it: pitch movement, timing, pauses, loudness and stress placement. Tone of voice usually also implies the emotion those features convey, which is an interpretation rather than a measurement.

Why does clip length matter so much?

Because prosody is a property of an utterance rather than of a moment. A few seconds may contain only one or two phrase boundaries, which is not enough to say anything about how the speaker’s timing and intonation are distributed. Longer ordinary-quality audio beats a short clean excerpt.