Most of the existing works on Indian transliteration use character mappings coupled with a set of rules to transliterate. The statistical approaches have hardly been tried with a few exceptions. Apart from few statistical approaches, which use phrase based SMT and generative joint source-channel model for transliteration, most of the works on transliteration use rules defined over character mappings. In this proposal, I propose statistical learning for transliteration. The model parameters are learned from transliteration pairs which are automatically mined from parallel corpora.