Tokyo, November 28, 2016 - NEC Corporation (NEC; TSE: 6701) today announced the development of an acoustic situation awareness technology that recognizes events and situations based on the corresponding sounds they produce.

Sound uniquely enables the recognition of situations that cannot be detected visually, such as events occurring on the opposite side of walls or crowds of people. Development efforts have long been underway to produce reliable sound recognition technologies, however, when targeting a wide area, the degradation in sound recognition due to other sounds (ambient environmental noise) has been a challenging issue.

NEC's new acoustic situation awareness technology, developed as part of its lineup of cutting-edge artificial intelligence (AI) technologies, NEC the WISE (*1), uses a combination of technologies to recognize situations based on the detection of multiple acoustic events (*2). For example, it uses constituent sound extraction technology to divide sounds collected by a microphone into target sounds and ambient noise, then extracts fine constituent sounds from the target sounds. It also uses event classification technology to discern the occurrence (or non-occurrence) of an event based on the combination patterns of constituent sounds.

"With this new acoustic situation awareness technology, NEC aims to quickly and accurately recognize dangerous situations in a variety of environments," said Akio Yamada, General Manager, Data Science Research Laboratories, NEC Corporation. "Not only can this technology contribute to the safety of our communities through the detection of sounds related to crimes, it can also be applied at the home to help alert family members of unexpected accidents."

The new technology won first place in the IEEE AASP Challenge: Detection and Classification of Acoustic Scenes and Events 2016 (DCASE2016), Task2 - Sound event detection in synthetic audio (*3), an international acoustic detection contest. NEC also confirmed, through in-house validation testing, that the new technology is capable of sufficiently detecting sound events at long distances up to five times the distances possible with existing technologies (*4).

The main features of the new technology are as follows:

• Constituent sound extraction technology, capable of high-sensitivity detection of even quiet sounds For example, the sound produced by glass breaking varies depending on the environment. It could be a loud crashing sound, or more of a ringing sound. By dividing the sounds collected by a microphone into fine constituent sounds (such as pre-learned component sounds) which are not affected by environmental differences and ambient environmental noise (unknown sounds which have not been learned by the AI), constituent sound extraction technology achieves the high-precision extraction of constituent sounds without being affected by environmental noise.

• Event classification technology that enables determination of events taking place Event classification technology discerns the occurrence (or non-occurrence) of target sounds by learning event patterns (combinations of constituent sounds that are unaffected by environmental differences) in advance and cross-referencing them with the high precision constituent sounds extracted by the constituent sound extraction technology.

The combination of these two technologies enables highly sensitive detection of sound events over wide areas in which large amounts of ambient environmental noise exist, without missing quiet sounds, and facilitates easy deployment into unknown environments without the need to teach the AI the target sounds on an environment-by-environment basis.

Note:

• (*1)

July 19, 2016 NEC announces new AI technology brand, "NEC the WISE"

http://www.nec.com/en/press/201607/global_20160719_01.html

NEC's AI (Artificial Intelligence) Research http://www.nec.com/en/global/rd/crl/ai/index.html

• (*2) The technology is not capable of discerning the meaning of conversations.

• (*3) URL: http://www.cs.tut.fi/sgn/arg/dcase2016/

• (*4) In validation testing, NEC achieved detection performance equivalent to that achieved by existing technologies at a range of 4m at a range of 20m, and confirmed performance by simulating the ability to achieve total coverage between surveillance camera installation positions.