With the popularitzation of music services on the internet, such as Spotify, Last.fm, and Tidal, systems which are able to recommmend songs based on user tastes have become necessary. Many such systems can be found in the literature, like tag based systems, signal based systems and social contribution systems. This paper presents a new tag based system that uses an LDA model as a dimensionality reduction tool, while mantaining quality of representation. The system is able to generate both random and deterministic playlists using simple distance metrics, with low complexity. The recommender quality is indirectly assessed, since A/B tests are costly and time consuming to evaluate such systems. For that reason we adopt a score based on the homogeneity of the playlists created by the system related to genre classification. We also propose new methods to deal with songs without tags, so that we can take advantage of the newly added songs in the dataset and deal with the sparsity and cold-start problems, inherent to with other recommendation systems. While the treatment of songs without tags requires further investigation to achieve acceptable results, the results indicated for the generated playlists by the homogeneity scores were very good. The proposed quality measure was able to shed a new light on the observation of resulting playlists.