Marmoset vocalizations are built from discrete vocal motor elements.

For decades, primate vocalizations have been cast as fixed action patterns - stereotyped, ballistic motor sequences that run to completion once initiated. Yet accumulating evidence suggests that this view is too coarse. We previously identified discrete, macro-scale, vocal subunits of ∼80 ms within the common marmoset phee call, revealed by noise-evoked cancellations1 and natural segmentation2. It remains unclear, however, whether this modularity is a special property of such long-distance calls or a fundamental principle across marmoset call types. By dissecting fine-grained temporal and spectral transitions across call types3, we find that vocalizations are constructed from a shared pool of discrete, rhythmically organized vocal motor elements (VMEs) in the gamma range (30-90 Hz4), shorter than vocal subunits in phee calls. VMEs follow a quantized logic: their durations and frequency slopes are conserved across call types, akin to a shared 'alphabet'. We propose a modular motor architecture model in which marmoset vocalizations are governed by a hierarchically nested, rhythmically gated set of VMEs. This provides a missing link between peripheral biophysical constraints and vocal evolution, suggesting that the combinatorial foundations of human speech may have emerged from pre-existing, rhythmically gated motor modules.