Generated singing has become convincing across short phrases, and it still gives itself away when a note is held. The failure is specific and points at what these systems model well and what they do not.
A held human note is never steady
A singer sustaining a pitch produces continuous small fluctuations in frequency and volume, driven by breath pressure and muscular control that cannot be held perfectly still.
Vibrato is the largest of these, and beneath it sit irregular micro-variations that no performer produces identically twice.
Listeners do not consciously hear these, and they hear their absence immediately as something mechanical.
Models are pulled toward the average
A system trained to predict the most likely continuation of an audio signal will favour the centre of the distribution.
Over a short phrase, the phonetic content and pitch movement carry the listener and there is little room for the average to assert itself.
Over a long note there is nothing but sustain, so the model outputs something close to a mean, and the mean is smoother than any real performance.
Vibrato added afterwards sounds applied
The obvious fix is to impose a periodic modulation on the sustained pitch, and it is the reason many generated vocals sound synthetic in a distinctive way.
Real vibrato varies in rate and depth across a single note, usually beginning late and widening as the note continues.
A regular oscillation of fixed rate and depth is detectable within a second, because human hearing is unusually sensitive to periodicity in pitch.
Breath does not run out
A singer holding a note is spending air, and the tone changes as pressure falls, often thinning or tightening toward the end of a phrase.
A generated voice has no such constraint and can sustain indefinitely without any of the accompanying change in timbre.
The result reads as effortless in a way that removes the sense of physical effort listeners associate with a demanding line.
The training data is short by construction
Audio models are typically trained on brief segments, because processing cost rises steeply with length.
Long sustained notes are comparatively rare in any corpus and are frequently split across segment boundaries, so the model sees fewer complete examples of exactly the case it handles worst.
This is why producers working with these tools tend to write around the limitation, favouring rhythmic and syllabic lines over exposed sustained ones.