The Computer Learns to Hear

At a noisy party a person is talking to a friend, glasses clink all around, music plays, twenty voices talk at once, and suddenly from the other end of the room he hears his own name. All evening his ear and brain had been doing work he did not notice: splitting the general hum into separate voices, following one of them and not losing track of the rest. In 1953 the British scientist Colin Cherry called this ability the cocktail party problem, and machines spent the next seventy years solving it.

The first machines that listened in place of a person were built for surveillance. In a sharashka Leon Theremin built the Buran, which reconstructed a conversation in a room from the vibration of a window pane, and his Zlatoust, hidden in a wooden coat of arms, spent seven years relaying to Soviet intelligence the conversations in the American ambassador’s office. These devices could hear, but working out who said what still fell to people in headphones.