From the paper, the IST on the graph stands for "Initial Speech Transcription", and those data points are for the speech-input text before any corrections were made. The other "Speech" data-points include time to make corrections either by using the keyboard or using speech recognition.