In this paper, we introduce a new approach to the computer transcription of handwritten Pitman shorthand as a rapid means of text entry (up to 100 words per minute) into today’s handheld devices, almost at the rate of speech. It is different from previous applications of the same framework from two aspects: – firstly, a novel idea of using geometric attributes other than phonetic attributes in the abstraction of a phonetic Pitman’s shorthand lexicon is proposed. Secondly, a Bayesian network representation for the organisation of shorthand-outline models is introduced, in which natural variability of Pitman shorthand is defined via different nodes and links. Using a probabilistic Bayesian network, the system shows a noticeable robustness not only in transcribing a variety of genuine handwriting, but also in estimating missing vowel components that may have been omitted in speed writing. The accuracy of the new approach (92.86%) is a considerable improvement over previous applications.