Nascent RNA sequencing protocols, such as GRO-seq and PRO-seq, are now widely used in the study of eukaryotic transcription, and these experimental techniques have given rise to a variety of statistical and machine-learning methods for data analysis. These computational methods, however, are generally designed to address specialized signal-processing or prediction tasks, rather than directly describing the dynamics of RNA polymerases as they move along the DNA template. Here, I introduce a general probabilistic model that describes the kinetics of transcription initiation, elongation, pause release, and termination, as well as the generation of sequencing read counts. I show that this generative model enables estimation of separate rates of initiation, pause-release, and termination, up to a proportionality constant. Furthermore, if applied to time-course data in a nonequilibrium setting, the model can be used to estimate elongation rates. This model additionally leads naturally to likelihood ratio tests for differences between genes, conditions, or species in various rates of interest. A version of the model in which read counts are assumed to be Poisson-distributed leads to convenient, closed-form solutions for parameter estimates and likelihood ratio tests. I present extensions to Bayesian inference and to a generalized linear model that can be used to discover genomic features associated with rates of elongation. Finally, I address technicalities concerning estimation of library size, normalization and sequencing replicates. Altogether, this modeling framework enables a unified treatment of many common tasks in the analysis of nascent RNA sequencing data.