Design Of Approximate Computing Module For Low Power Keyword Recognition

Keyword recognition techniques are useful to achieve better human-computer interaction,but in a low signalto-noise ratio environment,keyword recognition performance is currently difficult to meet the ideal requirements.In the face of more word classification tasks,traditional or mainstream neural networks also cannot meet the ideal requirements as well.From the algorithm level to the hardware level,this thesis implements an adaptive multikeyword speech recognition system that can handle a variety of noise environments.At the algorithm level,a convolutional neural network based on the SNR prediction module is implemented.Firstly,the SNR prediction module is introduced to predict environmental noise.In a low signal-to-noise ratio environment,a high-precision convolutional neural network based on an approximate multiplication unit with an 8-bit width is implemented.Secondly,in a high signal-to-noise ratio environment,in order to reduce the complexity of the neural network,non-uniform quantization method of the Deepshift neural network is used,which greatly reduces the calculation.Thirdly,in a low signal-to-noise ratio environment,in order to retain the accuracy of the neural network,a neural network with 8-bit weights whose performance is comparable to that of a full-precision network is used.When the signal-to-noise ratio is 0d B,the recognition result of 12 keywords is 80%.In the case of clean,the keyword recognition accuracy is over 88%.At the hardware level,the circuit design according to the designed algorithm is optimized to reduce power consumption.Firstly,circuit reconfigurable techniques are used in this thesis,so that CNN and Deepshift operation units can use partially shared modules.Secondly,the design of the shared modules is optimized to reduce the calculation and the circuit complexity.Thirdly,a separate approximate multiplication unit is designed for the multiplication and addition operations in the convolutional neural network,and the approximate booth coding method and approximate adder design is introduced.While ensuring the accuracy,the power consumption of the multiplier and the adder is reduced at the same time.