Timing is a critical component of a wide range of sensorimotor tasks that can span from a few milliseconds up to several minutes. While it is assumed that there exist several distributed systems that are dedicated for production and perception [1], the neuronal mechanisms underlying precise timing remain unclear. Here, we are interested in the neural mechanisms of sub-second timing with millisecond precision. To this end, we study the control of song timing in male Zebra Finches whose song production relies on the tight coordination of vocal muscles. There, the premotor nucleus HVC (proper name) is responsible for the precise control of timing. Current models of HVC rely on the synfire chain, a pure feed-forward network. However, synfire chains are fragile regarding noise and are only functional for a narrow range of feed-forward weights, requiring fine tuning during learning. In the present work, we propose that HVC can be modelled using a ring attractor model [2], where recurrent connections allow the formation of an activity bump that remains stable across a wide range of weights and different levels of noise. In the case of asymmetrical connectivity, the bump of activity can "move" across the network, hence providing precise timing. We explore the plasticity of syllable duration in this framework using a reward-driven learning paradigm and a reward-modulated covariance learning rule applied to the network's synaptic weights [3]. We show that the change in duration induced by the learning paradigm is specific to the target syllable, consistent with experimental data.