Amazon patent describes speech recognition prior to hearing Alexa wakeword
Rather than listen for a “wakeword,” a new patent application from Amazon suggests Alexa could simply record all speech, allowing users to give instructions without exactly following the prescribed order of uttering “Alexa” first, Business Insider reports.
The patent for “Pre-wakeword speech processing,” published recently by U.S. Patent and Trademark Office, describes a system in which speech is recorded, possibly in 10 to 30 second intervals, and processed for wakeword detection, possibly before being purged, and the next interval recorded.
“The [proposed] system is configured to capture speech that precedes and/or follows a wakeword, such that the speech associated with the command and wakeword can be included together and considered part of a single utterance that may be processed by a system,” the inventors write in the application.
Having a constant recording, instead of the “big.LITTLE” style architecture of wakewords, however, removes one level of assurance that Amazon is not recording every conversation for its own purposes, such as speech recognition training and annotation.
An Amazon spokesperson provided a standard statement to Engadget about the highly speculative nature of patent applications.
Sensory recently launched a platform for speech recognition on edge devices to preserve user privacy.