Extremely Low Power Keyword Spotting Neural Network Circuit Design

Voice wake-up technology is one of the important entrances for human-computer interaction.The wake-up function is usually realized by spotting a small number of keywords.Therefore,Keyword Spotting(KWS)has become a key technology in the field of speech recognition.The KWS circuit is usually used as a switch in an electronic system.This feature determines that the KWS circuit will be in an always-on state that is always running.Therefore,the extremely low power KWS circuit is demanded more and more strongly especially in the applications with limited battery capacity,such as the Internet of Things and wearable devices.An extremely low power end-to-end KWS circuit is designed based on deep neural network.The algorithm and model that implemented by KWS circuit are compressed and quantized by low precision to reduce the storage and computing resources of the circuit.A step-by-step hierarchical hybrid precision quantization method is proposed based on Mel Frequency Cepstrum Coefficient(MFCC)feature extraction algorithm.A lightweight deep separable convolution network with hybrid precision for two keyword recognition is trained and obtained with the methods such as binarization and fixed precision quantization.An extremely low power KWS circuit is proposed to implement the quantized feature extraction algorithm and the compressed neural network model.The main work and innovation of circuit design are as follows:(1)Feature extraction circuit is divided into sub-modules and pipelined frame-by-frame according to the operation steps,which reduces the operating frequency of feature extraction circuit to 1/4 of the original.(2)A neural network processor circuit is designed which is composed of a Process Element(PE)array,memory,data mapping module and a state machine.The PE unit adopts the hybrid precision Multiply and Accumulate(MAC)design.The data representation method is specially designed to reduce the power consumption of the PE array.(3)A noval data reuse method is proposed based on the input characteristics of the speech feature stream data,which reduces the 8-bit and 1-bit calculation amount of the network to 5% and 7.34% and storage of the intermediate data to 28.6% before reuse,respectively.The circuit is simulated using TSMC 28 nm CMOS process.Experiment shows that the KWS ciucuit proposed in this thesis recognizes 2 keywords with an accuracy rate of 95.6%.The circuit operates at a low frequency of 40 kHz.The power consumption of the feature extraction circuit is 0.28?W and the neural network processor is 0.12?W.The whole KWS circuit consumes only 0.4?W,which is significantly reduced compared to other similar designs.