MFCC vs aMFCC

With the recent development of IoT and mobile devices, the need for efficient voice-activated functions is significantly increasing. The traditional approach leverages cloud connectivity to execute speech recognition and natural language processing algorithms on powerful remote servers. While this method is effective, it can be challenging for applications such as battery-powered devices due to the significant energy consumption required to continuously stream audio data to the cloud.
Another approach would be to process wake-up keywords and a restricted set of command words on the end device itself, thereby reducing the power budget to a minimum. This will also allow to reduce the latency.
The analog MFCC (aMFCC) approach allows to manage this challenge by drastically reducing the power needed to extract the audio features commonly used by voice recognition algorithms.
This application note presents both the conventional MFCC and the aMFCC approaches and provides the key insights to understand their differences.