Background: Predicting aphasia recovery after stroke has been difficult due to substantial variability in outcomes. Few studies have characterised the nature and extent of recovery, beginning with baselines at 24-72 hours after stroke onset. Aim: To characterise the course of language recovery after first-time stroke. Methods: Using our Performance and Recovery in Stroke Study (PARIS) database, we evaluated consecutive first-time stroke patients with aphasia and diffusion-weighted-image-positive lesions on admission and at 90 days. Results: Twenty-two of 91 patients had language disorders. Initial syndrome scores were positively correlated with 90-day scores (r = 0.60) and negatively correlated with the change in score from baseline to follow-up (r= -0.66). Neither lesion size, age nor education correlated with initial syndrome severity or with performance at 90 days. Level of education was not associated with degree of recovery. A multiple regression model that combined lesion size, age and initial syndrome was significant (p = 0.03) but only explained 29% of the variance. Patients with severe deficits at baseline in individual language domains could recover, improve to a less severe deficit or not improve at all. Conclusion: There was significant variability in language recovery after first-time stroke, even in more severe, initial syndromes. Traditional predictors of post-stroke language outcomes did not reliably predict function at 90 days. These data suggest that other factors that account for functional stroke recovery have not yet been identified.