One of the key points in music recommendation is authoring engaging playlists\naccording to sentiment and emotions. While previous works were mostly based on\naudio for music discovery and playlists generation, we take advantage of our\nsynchronized lyrics dataset to combine text representations and music features\nin a novel way; we therefore introduce the Synchronized Lyrics Emotion Dataset.\nUnlike other approaches that randomly exploited the audio samples and the whole\ntext, our data is split according to the temporal information provided by the\nsynchronization between lyrics and audio. This work shows a comparison between\ntext-based and audio-based deep learning classification models using different\ntechniques from Natural Language Processing and Music Information Retrieval\ndomains. From the experiments on audio we conclude that using vocals only,\ninstead of the whole audio data improves the overall performances of the audio\nclassifier. In the lyrics experiments we exploit the state-of-the-art word\nrepresentations applied to the main Deep Learning architectures available in\nliterature. In our benchmarks the results show how the Bilinear LSTM classifier\nwith Attention based on fastText word embedding performs better than the CNN\napplied on audio.\n