Twitter data have brought new opportunities to know what happens in the world in real-time, and conduct studies on the human subjectivity on a diversity of issues and topics at large scale, which would not be feasible using traditional methods. However, as well as these data represent a valuable source, a vast amount of noise can be found in them. Because of the brevity of texts and the widespread use of mobile devices, non-standard word forms abound in tweets, which degrade the performance of Natural Language Processing tools. In this paper, a lexical normalization system of tweets written in Spanish is presented. The system suggests normalization candidates for out-of-vocabulary (OOV) words based on similarity of graphemes or phonemes. Using contextual information, the best correction candidate for a word is selected. Experimental results show that the system correctly detects OOV words and the most of cases suggests the proper corrections. Together with this, results indicate a room for improvement in the correction candidate selection. Compared with other methods, the overall performance of the system is above-average and competitive to different approaches in the literature.