Listening to an AI voiceover with measurements, not impressions

A synthetic read fails in three specific ways

Generated voice has become good enough that the first reaction to it is usually approval. Then a listener says, without being able to explain, that it sounds like an advert. The vagueness is the problem: a note that says it sounds off cannot be acted on, and a note that says the pitch never moves can.

Three traits cover most of the dissatisfaction. Flat pitch: the fundamental frequency stays within a narrow band, so every sentence carries the same emphasis. No breath: real speakers pause for air, and a synthetic read often runs sentences together with gaps too short for a person to breathe. Uniform cadence: the rhythm is metronomic, with every clause lasting about as long as its neighbour.

Each of those is measurable from the waveform, and each measurement is more useful than the adjective that prompted it. You can tell the studio that the pitch spans 34 hertz rather than that it sounds flat, and that is a note which produces a different take.

What the waveform actually shows

The waveform gives you the shape of loudness over time. Speech sits between pauses; a long flat stretch is either silence or a held note, and a wildly varying block is usually a plosive or a level change that was edited rather than performed.

The three figures below it summarise the traits. Pitch range is the spread of detected fundamental frequency across voiced speech — low numbers mean monotone, very high numbers often mean a detector confused by noise. The pause figure counts gaps longer than 200 milliseconds and reports the longest one, which is the breath a listener unconsciously waits for.

Cadence is the third: how much sentence length varies across the take. A read with no variation sounds like a list being read out; a read with too much sounds unstable. The measurement does not decide which one you have, but it makes the difference visible between two takes that both sound plausible on a first listen.

Reading the numbers against the audio is the point. Play the take, watch where the waveform peaks, then jump to the moment the pitch figure drops. The measurement tells you where to listen; the listening tells you whether the number describes a problem.

• Pitch range: narrow means monotone, implausibly wide means measurement noise.
• Pause count and longest pause: a breath, or the absence of one.
• Cadence variation: metronomic or unstable.
• Peaks and clipping: a level problem no take can fix in the mix.

The order to judge in

Start with intelligibility. A voiceover that cannot be understood at normal listening volume has no other qualities worth discussing, and no amount of pitch variation fixes a swallowed consonant. Listen once on headphones, then once on a phone speaker, because a mix that works on one regularly fails on the other.

Then prosody, which is what the measurements describe. Does the pitch move on the stressed word in each sentence? Does the sentence end lower than it began, or trail upward into a question mark? Those are the traits that make a read sound like a person making a statement.

Then pacing against the picture, if there is one. A read can be beautifully varied and still be thirty seconds too long for its footage; the timing problem is usually solved in the script rather than in the synthesis settings.

Then the room. Breath gaps, mouth clicks and the small acoustics of a space are the details that decide whether a listener believes they are being spoken to. Synthetic takes tend to be too clean in a way that reads as intimate and wrong, and a touch of room tone fixes more of it than people expect.

Settings versus performance

Most flat reads are not a limitation of the voice, they are a setting. Stability, style and similarity sliders trade consistency against variation, and the default sits where the output is safe: same pitch, same pace, no surprises. Moving one notch produces a different take and often a better one.

The measurements tell you which way to move. A narrow pitch range wants less stability or more style. A take with no breath wants the sentence segmentation broken up in the script, because a synthesiser only pauses where the punctuation gives it permission.

Cadence is the hardest to fix with settings, because it is mostly a property of the writing. Sentences of similar length produce a metronome whatever the slider says, and the fastest fix is to rewrite the two longest sentences into four.

Listen to three takes rather than one and compare the figures before choosing. The difference between takes is usually bigger than the difference between settings, and a take chosen by ear alone tends to be the one where the voice happened to land the first sentence.

Reading the take against the script

The waveform tells you how the read behaved; the script tells you whether it behaved in the right places. Line the two up: the emphasis should fall on the word that carries the meaning in each sentence, and the pause should fall where a listener needs a moment.

A common failure is an emphasised number. Synthesisers will stress a figure because it is syntactically prominent, and stress on a number in a sentence about cost reads as a sales pitch. Rewriting the sentence so the number is a dependent clause usually fixes it without touching the voice.

Check the pronunciation of names, products and acronyms against the written form. Every synthesiser mispronounces something, and the ones it gets wrong are the ones your client cares about. Build a list of those words and their phonetic spellings, and reuse it across projects.

Finally, listen once at the volume the audience will use. A read that sounds warm on headphones can sound thin on a phone speaker, and a take that passes every measurement can still fail the only test that matters, which is whether a person would keep listening.

The script is the performance

Most of what people dislike about a synthetic voice is in the writing. Long sentences with three subordinate clauses give a synthesiser no place to breathe, and a paragraph of identically structured sentences produces the metronome no slider can remove.

Write for the ear: one idea per sentence, a full stop where the voice should pause, and the stressed word placed early enough to be heard as emphasis rather than as an accident. Read the script aloud yourself before generating it, and mark where you breathe.

Then use the measurements to test the writing rather than the voice. A take with a narrow pitch range is often a script with no contrast: no question, no list, no comparison, nothing for the voice to push against.

This is also why a voiceover review belongs before the soundtrack, not after. Changing the script costs nothing; re-cutting a mix costs a day.

Two takes, one decision

Comparing two takes is more useful than judging one. The figures give you a difference rather than a value: this take spans forty hertz more than the last, this one has three long pauses where the other had none.

That difference is what you can act on, because it corresponds to a change in settings or in the script. A single number only tells you whether the take is near some threshold that nobody agreed on.

Where both takes measure the same, the choice is a judgement about tone, and that is the point at which a person should listen rather than a chart should decide.

What no measurement can tell you

Whether the voice suits the brand, the audience or the moment. A warm, slow read is right for a memorial video and wrong for a product launch, and both may show identical figures.

Whether the script is any good. A perfectly delivered line can be factually wrong, and a voiceover is the last place in a pipeline where a claim gets spoken without a source attached. Everything a listener hears as fact should have been checked before it was recorded.

Whether the voice is licensed for the use you have in mind. Synthetic voices carry terms that vary by platform, by commercial context and by whether the output is used in advertising rather than content. That is a paperwork question, and it is the one most often skipped.

Whether the person being imitated consented, if the voice resembles someone real. No technical check answers that, and it is the question that ends projects.