Abstract: Methods, systems and apparatuses are provided for speech recognition using a touch prediction model for a multi-modal system which include: pre-training the touch prediction model of the multi-model system to enable a prediction of probable commands of a subsequent user based on a trained model wherein the trained model is trained using a history of a set of touch actions by plurality of users; sending at least parameter data associated with a current page displayed to a particular user at an instance of interaction with the multi-model system by the particular user; receiving inputs from a group including: previous touch actions, system parameters and contextual parameters of the plurality of users; and using a set of options associated with the top n-most probable commands upon receipt of speech commands by the user by using a vocabulary associated with a reduced set of probable commands for recognition.