To understand the speech of someone amidst a crowd of many people, almost no effort is required. So far this has been impossible for machines.
Japanese company Mitsubishi Electric introduced a new technology for recognizing human speech. AI successfully divides and recognizes the voices of one of two people, simultaneously speaking into the same microphone but in different languages. Recognition is almost real-time, with a maximum delay of 3 seconds. Accuracy is 90%, and when three people speak at the same time, the accuracy drops to 80%. High accuracy is achieved using a variety of Deep Learning. The neural network divides the speech into small particles, and a special algorithm selects common elements.
An important feature of the new speech recognition system is that it does not require pre-training or any preparation to work with a person's voice and can work with any voices without pre-training. Accuracy does not depend on the language. "We will use the new technology in places where high recognition accuracy, such as driving new cars, elevators, household appliances and other electronic devices, is needed," said Yohei Okato, a Mitsubishi Electric representative.
Sources: 1; 2; 3; 4
-
