Voice recognition traditionally works by dividing the words you speak into smaller parts known as phonemes, according to Science Line . The speech recognizer will then analyze the order, template, and context of phonemes to generate a sentence. This usually results in a delay while you wait for the software to predict exactly what you are saying. The Google AI team said in the publication that its model "takes characters one by one while you speak, with white spaces where appropriate". In other words, Gboard will recognize separate letters from the words you say, as you say, instead of recognizing phonemes.