Combining word prediction and r-ary Huffman coding for text entry

Two approaches to reducing effort in switch-based text entry for augmentative and alternative communication devices are word prediction and efficient coding schemes, such as Huffman. However, character distributions that inform the latter have never accounted for the use of the former. In this paper, we provide the first combination of Huffman codes and word prediction, using both trigram and long short term memory (LSTM) language models. Results show a significant effect of the length of word prediction lists, and up to 41.46% switch-stroke savings using a trigram model.