Spectrograms from 14,759 short voice clips gave researchers a balanced accuracy near 72% for spotting diabetes, using sound alone.