While temporal control is crucial for the generation of a wide range of sensorimotor tasks, its underlying mechanisms remain unclear. Like human speech, birdsong relies on a tight muscle coordination, and songbirds have proved to be an outstanding model to study the sequential pattern of neuronal activity encoding action timing. In male zebra finches, the premotor nucleus HVC is responsible for the precise control of song tempo. Current computational models of HVC rely on synfire chains, a purely feedforward network model that can account for HVC sequential activity. Synfire chains are however not robust to noise and function for a narrow range of feedforward weights, thus requiring fine tuning during learning. On the contrary, attractor dynamics provide networks with robust functional properties that make them an attractive alternative to feedforward models. Here, we propose that HVC neuronal dynamics may be modelled using the ring model (Zhang, 1996; Hansel & Sompolinsky, 1998), where recurrent connections allow the formation of an activity bump that remains stable across a wide range of weights. In the case of asymmetrical connectivity, the bump of activity moves across the network. We show that the width of the activity bump, and thus the duration of transient neuronal activation, can be decreased to reproduce the brief activity bursts of HVC neurons. Relying on a reward covariance rule (Williams, 1992), we show that the duration of a syllable can be modified in response to a perturbed reward profile, as implemented in a widely used lab conditioning paradigm. Consistent with behavioral data, the change in duration is specific to the target syllable. We derive a new prediction from the model and show that following local muscimol (GABAa agonist) injection in HVC, a perturbation in the initial phase of the pattern formation would delay song initiation, but singing would be sustained.