Hello Edge: Keyword Spotting on Microcontrollers

Sub-mW Keyword Spotting on an MCU: Analog Binary Feature Extraction and Binary Neural Networks

“…Previous approaches relying on the conventional ADC + preprocessing flow such as Hello Edge [24] and EGRU, [31] even with an extremely optimized DNN model or hardware accelerator, will have an upper limit on the energy efficiency due to the data acquisition and pre-processing overhead. Our proposed system architecture with a binarizing AFE provides a method to overcome this fundamental limitation.…”

“…On their left, there are the systems that include the analog front end and binarization of weights ad activations. Here, the system achieves sub-mW consumption with 80% accuracy in the task, 4% less accuracy than the solution from HelloEdge [24], which consumes more than 10 mW.…”

“…Unfortunately, activations are more sensitive to quantization in terms of accuracy drop. The approximation of the activations is quasi-lossless down to 8 bit [24], and it allows the hardware to parallelize four computations over a 32-bit register.…”

“…Pioneers of edge computing for keyword spotting [24] manage to implement a keyword spotting model on a cortex M4 based MCU: they explore a different kind of models, and they reach the state of the art accuracy even in memoryconstrained devices. Moreover, they present an implementation with 8-bit quantization for weights and activations to use SIMD instruction and parallelize four operations in just one instruction.…”

“…In the proposed system, the same microphone is connected to all these filters. The next step is to choose the right filter shape for further classification: log-Mel spectrogram and the following MFCC features are intensively used in sound-based classification systems [13], [24]. For this reason, we choose corner frequencies equally-spaced in the Mel domain within the selected range.…”