Neural Network-Based Speech Keyword Recognition Algorithm And Circuit Design For Low Signal-To-Noise Ratio

Voice keyword recognition helps to achieve better human-computer interaction.From algorithm to hardware,this thesis implements a keyword recognition scheme that can be applied to low signal-to-noise environment.Algorithmically,an end-to-end speech keyword recognition network is realized.First,in this thesis,for speech signal recognition with low signal-to-noise ratio,the traditional feature extraction algorithm is replaced by a single-dimensional convolutional neural network to avoid the performance loss of manual feature extraction algorithm under low signal-to-noise ratio.Second,this article uses a fixed-point quantization training method to quantize the weight and data to 8 bits.When the signal-to-noise ratio is-5d B,the recognition accuracy of the four keywords can be up to 87%,and the accuracy rate of mixed speech keyword recognition with the signal-to-noise ratio of-5d B to20d B is 87.3%.On the hardware,the circuit design is optimized for the algorithm to reduce power consumption.First,the use of circuit reconfiguration enables centralized implementation of convolution and fully connected operations,avoids the computing power loss caused by the distributed architecture,and increases the equivalent computing power of each multiplier by 36%.Second,the hardware uses a 16-segment linear iterative approximation scheme for the nonlinear activation function,which greatly reduces the complexity of the circuit.Third,the multiplication and addition operation of the neural network adopts the design of approximate multiplication and approximate adder based on bit pair encoding,which let the power consumption of the multiplier and adder is reduced by 30%while ensuring accuracy.The circuit design of this paper adopts TSMC 22nm process,and the area after layout and wiring is 0.76mm~2.When the working frequency is 4MHz,real-time processing of speech can be realized,and the power consumption is about 7.4?W.