Here’s an important development—TechCrunch reports, “MIT Develops a Speech Recognition Chip that Uses a Fraction of the Power of Existing Technologies.” This could lead to hardware that stays charged for months at a time. MIT professor Anantha Chandrakasan notes his team’s breakthrough should benefit the wearable tech field:
The miniaturization of these devices will require a different interface …. It will be critical to embed the speech functionality locally to save system energy consumption compared to performing this operation in the cloud.
The chip can distinguish background noise from speech, and switches on when it detects the latter (so power is always “on” a tiny bit). The team used the open source Kaldi toolkit to train the speech recognition neural network on several datasets.
Cynthia Murrell, March 9, 2017