For improving the accuracy of a speech recognition system, for the specific task of speech-to-text (dictation style speech) translation, a constrained dictation methodology using speak-and-spell mode is disclosed. The invention is perfectly suited for modern day “text-messaging” applications wherein the number of words being dictated is very small (limited by the 140-160 characters message length constraint). Additionally, the invention adds a control on the way users interact with machines, thereby making the speech recognition task easier and improving system accuracy.