Photo: Google
Google Assistant is a key part of the experience with Android Auto, but as many of us learned the hard way, the voice commands aren’t working exactly as anticipated in the car.
The search giant now wants to improve the “Hey Google” wake-up phrase detection model using a new approach, all based on federated learning.
In other words, Google wants to use voice recordings of you saying “Hey Google,” though worth noting is that the data never leaves your device. The company uses a very clever approach that performs the entire analysis on your device and then sends the results to Google, thus without including any recordings or private information.
“Federated learning uses the voice recordings stored on users’ devices to refine models like ‘Hey Google.’ It learns how to adjust the model from the voice data, and sends a summary of the model changes to Google servers. To provide a better model for everyone, these summaries are aggregated across users,” the company explains.
Needless to say, the federated learning feature is turned off by default, and it only performs an analysis of the recordings when the device is idle, plugged in, and connected to Wi-Fi to make sure no impact on the performance of your phone occurs.
Google explains in a support document that what it wants to improve is how the wake-up phrase is activated, as it determined that in some cases, it fails to launch Google Assistant even if you say “Hey Google.” In some instances, the assistant is launched without you even saying the wake-up words.
For drivers, all these improvements would certainly come in handy, especially because detecting the “Hey Google” phrase in the car isn’t as easy as in a quiet environment given the engine noise and the other sounds in the cabin. Time will tell how Google Assistant would improve based on federated learning, but on paper, this sounds like the proper way to go for “Hey Google” refinements.
In other words, Google wants to use voice recordings of you saying “Hey Google,” though worth noting is that the data never leaves your device. The company uses a very clever approach that performs the entire analysis on your device and then sends the results to Google, thus without including any recordings or private information.
“Federated learning uses the voice recordings stored on users’ devices to refine models like ‘Hey Google.’ It learns how to adjust the model from the voice data, and sends a summary of the model changes to Google servers. To provide a better model for everyone, these summaries are aggregated across users,” the company explains.
Needless to say, the federated learning feature is turned off by default, and it only performs an analysis of the recordings when the device is idle, plugged in, and connected to Wi-Fi to make sure no impact on the performance of your phone occurs.
Google explains in a support document that what it wants to improve is how the wake-up phrase is activated, as it determined that in some cases, it fails to launch Google Assistant even if you say “Hey Google.” In some instances, the assistant is launched without you even saying the wake-up words.
For drivers, all these improvements would certainly come in handy, especially because detecting the “Hey Google” phrase in the car isn’t as easy as in a quiet environment given the engine noise and the other sounds in the cabin. Time will tell how Google Assistant would improve based on federated learning, but on paper, this sounds like the proper way to go for “Hey Google” refinements.
Google News