Abstarct: In this thesis, a method for online Persian handwriting recognition baxsed on letter segmentation and recognition using HMM is proposed. Typed or handwritten image is called offline because its data is available after writing is completed whiles handwritten on digital devices like touch cell phones and tablets is called online since its data is available simultaneous with writing. Feature extraction from offline handwritten needs image processing but features of online handwritten are extracted directly from its information including horizontal and vertical co-ordinates. Sub-words were initially over-segmented baxsed on the vector connecting consecutive point’s angle with x-axis and then filtered for some false results using simple rules. Regarding existence of some extra segmentation points, each combination of sub-word’s segments is scored as the mean of HMM normalized probabilities of its segments and their top scores are chosen as the final candidates for letter group recognition, candidates which don't match any sub-word in lexicon are removed, then remained candidates which are not the same as input sub-word regarding existence of up and down signs are deleted. Then those candidates with non-identical number of strokes up and down to the input sub-word are deleted. Segmentation had accuracy of 85.47% and 60.1% of found segmentation points were false. Sub-word recognition was 37.85% for first candidate and 73.34% for first-10 candidates. This accuracy is very lower than holistic methods previously performed on the same databaxse. It's because of segmentation-baxsed method nature which is heavily dependent to segmentation stage and poor segmentation leads to low accuracy of sub-word recognition. No similar research is performed on the same databaxse yet for comparison. The advantage of segmentation-baxsed method over holistic method is that it can recognize each sub-word in the text dictionary, but the holistic method recognizes only handwritten samples available in the databaxse; in this method, however, segmentation stage errors extends to the recognition stage degrading dramatically the whole accuracy.