The attention mechanisms are playing a boosting role in advancements in\nsequence-to-sequence problems. Transformer architecture achieved new state of\nthe art results in machine translation, and it's variants are since being\nintroduced in several other sequence-to-sequence problems. Problems which\ninvolve a shared vocabulary, can benefit from the similar semantic and\nsyntactic structure in the source and target sentences. With the motivation of\nbuilding a reliable and fast post-processing textual module to assist all the\ntext-related use cases in mobile phones, we take on the popular spell\ncorrection problem. In this paper, we propose multi encoder-single decoder\nvariation of conventional transformer. Outputs from the three encoders with\ncharacter level 1-gram, 2-grams and 3-grams inputs are attended in hierarchical\nfashion in the decoder. The context vectors from the encoders clubbed with\nself-attention amplify the n-gram properties at the character level and helps\nin accurate decoding. We demonstrate our model on spell correction dataset from\nSamsung Research, and report significant improvement of 0.11\\%, 0.32\\% and\n0.69\\% in character (CER), word (WER) and sentence (SER) error rates from\nexisting state-of-the-art machine-translation architectures. Our architecture\nis also trains ~7.8 times faster, and is only about 1/3 in size from the next\nmost accurate model.\n