Voiceover waveform inspector

Robotic delivery is rarely one flaw. It is a pitch contour that barely moves, no pauses where a person would breathe, and every syllable spaced as evenly as a click track. Each one is measurable from the decoded samples, in the tab, without sending the audio anywhere.

Pitch comes from autocorrelation on the voiced part of the signal, decimated to 8 kHz so a minute of audio measures in about a second. The figure shown is the span between the low and high deciles in semitones, not the overall maximum, so one excited syllable does not make a flat read look dynamic.

A gap counts when the level stays under six percent of the loudest frame for more than 240 milliseconds, so a stop consonant does not register as a breath. Two numbers come out of it: pauses per minute and the longest pause in the take.

Cadence is the spacing between energy onsets, reported as the standard deviation divided by the mean. Under 0.15 the intervals are close to uniform, which is what a metronome sounds like.

These are measurements of a signal, not a verdict on a performance. A deliberately flat narration for a training module is a choice, and a slow read with long pauses is not more human by itself. What the numbers do is point at the moment to listen to: the longest gap, the flattest stretch.

Only the first 90 seconds are measured, and only the first channel. Nothing is uploaded: the file is decoded in this tab and discarded when the page closes.