Sensory Unveils Wake Word AI That Supports Multiple Voice Assistants
Sensory, a provider of embedded voice, vision, and biometrics technologies, has updated its TrulyHandsfree wake word engine with new support for multiple wake words from any combination of products, custom branded wake words, small vocabulary command sets, or even user-defined wake words.
This technology supports Amazon and the numerous supporters of the Voice Interoperability Initiative (VII) announced in September and allows companies to jumpstart development of products that feature multiple simultaneous voice services. "Multiple simultaneous wake words provide the best option for customers," Amazon founder and CEO Jeff Bezos said in Amazon’s VII press release.
"We are lucky to have worked closely with most of the leaders in voice services to help create products with their wake words embedded. Because of this, we already have all of the data in place to combine multiple models," said Todd Mozer, CEO of Sensory, in a statement. "However, everyone in the industry has found that combining two or more wake word models into a product poses a significant challenge in maintaining optimal performance. The new capabilities we added to TrulyHandsfree overcomes some of the accuracy challenges without significantly boosting MIPS and memory requirements."
Sensory has already built models for Microsoft's Cortana, Amazon's Alexa, OK Google, Apple's Siri, Samsung's Bixby, Baidu's Xiao-du, and more than two dozen other assistants. These wake words have been approved for market and are already shipping in real products like wearables, mobile phones, vehicles, smart speakers, and home appliances.
TrulyHandsfree employs a new technique to enable multiple wake words without linearly increasing the error rates, which has been the main challenge associated with running multiple wake words on one device. With traditional approaches, adding a new wake word meant the sum of the error rates of each individual wake word, but with TrulyHandsfree the accuracy doesn't degrade as fast, offering support for more wake words with less false accepts and false rejects.
"Also, on our roadmap for a 2020 release is multi-wake word adaptation that will allow devices to know who is talking to them and to adapt the voice model to better work with each individual user,: Mozer added.