In this paper, a novel model for synthesizing dance movements from music/audio sequence is proposed, which has variety of potential applications, e.g. virtual reality. For a given unheard song, in order to generate musically meaningful and natural dance movements, the following criteria should be met: 1) the rhythm between the dance action and music beat should be harmonious; 2) the generated dance movements should have notable and natural variations. Specifically, a sequence to sequence (Seq2Seq) learning architecture that leverages Long Short-Term Memory (LSTM) and Self-Attention mechanism (SA) is proposed for dance generation. The work in this article is interesting in the following aspects: 1) A cross-domain Seq2Seq learning framework is proposed for realistic dance generation; 2) A set of evaluation criterion is proposed for synthetization evaluation which do not have source for reference; 3) A dance dataset that including both music and corresponding dance motions collected, and very competitive results have been obtained against the-state-of-the-arts.