We present a novel system for automatic music mixing combining diverse music information retrieval (MIR) techniques and sources for song selection and transitioning. Specifically, we explore how music source separation and stem analysis can contribute to the task of music similarity calculation by modifying incompatible stems using a rule-based approach and investigate how audio-based similarity measures can be supplemented by lyrics as contextual information to capture more aspects of music. Additionally, we propose a novel approach for tempo detection, outperforming state-of-the-art techniques in low error-tolerance windows. We evaluate our approaches using a listening experiment and compare them to a state-of-the-art model as a baseline. The results show that our approach to automatic song selection and automated music mixing significantly outperforms the baseline and that our rule-based stem removal approach significantly enhances the perceived quality of a mix. No improvement can be observed for the inclusion of contextual information, i.e., mood information derived from lyrics, into the music similarity measure.