Tracking macroeconomic data at a high frequency is difficult as most time series are only available at a low frequency. Recently, the development of macroeconomic nowcasters to infer the current position of the economic cycle has attracted the attention of both academics and practitioners, with most of the central banks having developed statistical tools to track their economic situation. The specifications usually rely on a Markov-switching dynamic factor model with mixed-frequency data whose states allow for the identification of recession and expansion periods. However, such models are notoriously not robust to the occurrence of extreme shocks such as Covid-19. In this paper, we show how the addition of time-varying volatilities in the dynamics of the model alleviates the effect of extreme observations and renders the dating of recessions more robust. Both stochastic and conditional volatility models are considered and we adapt recent Bayesian estimation techniques to infer the competing models parameters. We illustrate the good behavior of our estimation procedure as well as the robustness of our proposed model to various misspecifications through simulations. Additionally, in a real data exercise, it is shown how, both insample and in an out-of-sample exercise, the inclusion of a dynamic volatility component is beneficial for the identification of phases of the US economy.