Automatic evaluation of verbal activities has always been hard. On the one hand, teachers want to stimulate student engagement and therefore often base grades on the number of participations. On the other hand, if quantity is more important than quality, students might participate, but add nothing new to the conversation. Providing feedback for spoken conversations is another challenge, one that is often solved by doing things manually. Both aspects originate from the same problem, the difficulty of extracting semantic knowledge from speech. In this contribution we present our approach to extracting and visualising this semantic knowledge using an open source speech recognition engine and Latent Semantic Analysis.