Despite the popularity of audio analytics, there are several challenges that hinder its adoption.
Despite the popularity of audio analytics, there are several challenges that hinder its adoption. According to Chris Mitchel, CEO and Founder of the UK-based Audio Analytic, voice recognition is technically a discrete problem.
“There are only a certain number of sounds the human mouth can make (phonemes) and the mathematical probability encoded in individual languages dictates the likely order these phonemes occur in,” Mitchel said. “To help software identify the relevant speech to decipher from a background of sounds, most voice assistants require the use of a trigger word - such as ‘Siri’, ‘Alexa’ or ‘hey, Google’ - to kick start the recognition.”
Sound recognition on the other hand is a far more complex technical problem to solve. Sounds can occur in any order, and in a far greater variety than the limited number of phonemes within human speech. Moreover, sounds are not preceded with a trigger word to help a system pick out what is relevant from the background mix of sounds. For example, an intruder gaining entry to a property is unlikely to pre-announce their activity by shouting “window break!”
“Training our software to recognize sounds requires advanced machine learning trained on actual real-world sounds,” Mitchel continued. “This means gathering millions of audio examples through a rigorous testing and research program. As part of this approach, we’ve broken windows of every size and type, sounded every available type of smoke alarm, gathered recordings of babies crying and dogs barking from real home environments. Our team are continually working on identifying, researching and mapping new sounds that each tell a significant story within the home.”
Richard Brent, CEO of Louroe Electronics added that there are two challenges that need to be overcome – first, know the environment that the detector is place into and understand the acoustics of that space. Second, be clear on the expectation of the beneficial performance.
According to Christian Connors, CEO of Shooter Detection Systems (SDS), audio analytic capabilities face two significant challenges: 1) Privacy laws and general concern of eavesdropping by the public; and 2) Historically high rates of false alerts causing the need for human monitoring and interpreting analytics.
“SDS has combined audio analytics with infrared signature analytics (dual-mode) to validate analytics without requiring streaming audio and potentially invading the privacy of the very individuals or entities that the system is meant to protect,” he said.