VLSI Design And Implementation Of High Precision Voice Activity Detection With Low Signal-to-noise Ratio

The end-point detection of human voice signals is becoming more and more widely used.Due to the complex and changeable scenes of the device,in some special scenes,the background noise will be very large.At this time,a variable noise environment is needed.In the low signal-to-noise ratio,there can be a detection circuit with higher recognition accuracy.Based on this,this thesis designs and implements a high-precision endpoint detection circuit for low signal-to-noise ratio and non-specific noise scenarios.This thesis proposes and adds a spectrogram feature enhancement module based on the basic algorithm framework for low signal-to-noise ratio voice activity detection,which further improves speech quality and recognition accuracy.Then the circuit design is optimized at three levels:parameter optimization,architecture optimization and unit optimization.Especially for the FFT module with the largest proportion of calculation in circuit design,a 512-point Fourier transform algorithm architecture combining radix 2 and radix 2~2 is proposed.And cascading pipeline multi-path delay cross structure is used to achieve the FFT module circuit,reducing the amount of calculation and greatly reducing the overall power consumption.This thesis implements and validates the algorithm model based on Matlab,uses TIMIT,HLDD,and Noise X-92 databases for testing,and implements it based on the TSMC 22nm CMOS ULP process.It is verified by VCS simulation and analyzed by DC synthesis and PTPX power consumption.In the non-specific noise environment of 0d B,-5d B and-10d B,they have accuracy rates of 94.56%,92.47%and 85.23%respectively,with a total power consumption of5.967?w and a total area of 1.55mm~2.Compared with similar designs,it can have higher recognition accuracy in non-specific noise type scenarios with low signal-to-noise ratio of 0d B,-5d B and-10d B,and the power consumption is also lower.