Social media networks and chatting platforms often use an informal version of\nnatural text. Adversarial spelling attacks also tend to alter the input text by\nmodifying the characters in the text. Normalizing these texts is an essential\nstep for various applications like language translation and text to speech\nsynthesis where the models are trained over clean regular English language. We\npropose a new robust model to perform text normalization.\n Our system uses the BERT language model to predict the masked words that\ncorrespond to the unnormalized words. We propose two unique masking strategies\nthat try to replace the unnormalized words in the text with their root form\nusing a unique score based on phonetic and string similarity metrics.We use\nhuman-centric evaluations where volunteers were asked to rank the normalized\ntext. Our strategies yield an accuracy of 86.7% and 83.2% which indicates the\neffectiveness of our system in dealing with text normalization.\n