Apple introduces machine learning technology to Siri so it no longer misidentifies the owner

Recently, Apple published its first article in its own machine learning journal "Apple Machine Learning Journal", which is mainly about the humanization process when users activate the "Hey Siri" function on iOS devices. This article also seems to mean that in the future, the iOS system will be equipped with a more advanced Siri intelligent personal assistant, at least able to use machine learning features to very accurately identify the owner of the device.

The "Hey Siri" feature was first introduced on the iPhone 6 with iOS 8. Now, on all Apple devices, the "Hey Siri" command can activate the smart personal assistant on the device. Of course, before the 6s, the device still needed to be plugged into a power source before it could be used, but on the current new iPhone and iPad, "Hey Siri" is now available 24/7, without the screen on or the power source being charged. Therefore, after the command is activated, users can follow up with inquiries or execute related voice commands, such as asking questions like "What's the weather like today?"

In this new article in the Machine Learning Journal, the Apple Siri team described the technical method of introducing the "Speaker Recognition System", claiming that in order to create a more humane device for users, the team built a "deep neural network", and this move will lay the foundation for the iterative update of Siri. In short, Apple has basically confirmed that in future versions of the iOS system, machine learning technology will be introduced into the Siri intelligent personal assistant to recognize the voice of the device owner.

Apple's Siri team said that they initially chose the phrase "Hey Siri" because it is a very "natural" wording. But it is also because of this that "Hey Siri" is often inadvertently activated, which brings a lot of trouble. Apple listed three situations, including: "When the user says a similar phrase", "When other users say 'Hey Siri'", and "When its users say similar phrases." According to the Siri team, the "false activation" in the first scenario is the most annoying.

In order to reduce the accidental and erroneous activation of Siri, Apple introduced technology from the field of speaker recognition research. The key point is that the Siri team said they are concerned about "who is speaking" rather than "what is said." Apple said that the goal of speaker recognition technology is to determine the identity of the person who speaks. "We are interested in 'who is speaking', not the problem of speech recognition, and the ultimate goal is to determine 'what is said'."

The Siri team said that other challenges currently faced by speaker recognition technology are how to achieve higher quality performance in reverberant (such as large rooms) and noisy environments. In the future, through machine learning, the Siri voice assistant will be trained to recognize the user's voice in a wide range of environments, such as in the car, kitchen, coffee shop or similar public places. Siri will also learn to recognize the state of the user's voice in various situations, including normal voice, high and bright voice, or the sound of groggy morning wake-up, etc., rather than being limited to the initial repeated recording of "Hey Siri".

Recently, the issue of Siri privacy loopholes has attracted much attention. Under this loophole, anyone can read the notifications of the device owner through Siri. Just hold down the HOME button or the side button on the iPhone X for a few seconds, call out Siri, and shout "Help me read the notifications", it will directly read the hidden message content on the lock screen. Whether it is a colleague, family member, or even a passerby, they can perform the same operation, and when the device is locked.

Obviously, Apple's solution of using machine learning technology to help Siri identify the owner of the device will solve the privacy issues faced by Siri.