Design Of Low-power Voice Activity Detection Module For Keyword Spotting System

Voice activity detection is a speech classification technology used to distinguish between voice and non-voice.It is often used as a switch for keyword spotting/speech recognition,so the recognition rate of speech activity detection is crucial for the operation of the post-level system;Voice activity detection is increasingly used in battery-powered IoT and wearable devices,and needs to be always on,so the motivation to reduce power consumption for voice activity detection is becoming stronger.Thus,a low-power voice activity detection module is designed,and the design is optimized from both the algorithm and circuitry aspects,so that it has both high recognition rate and low power consumption.Low power design is performed in both algorithm and circuit aspects.The neural network algorithm used for low-power voice activity detection is first compressed to reduce the consumption of computing resources and storage resources.Through the binarization and quantization,the bit-width of the data and weight of the neural network is compressed to obtain a binarized neural network(BNN)for the high signal-to-noise ratio(SNR)environment and a binarized weight network(BWN)for the low signal-to-noise ratio environment.Second,the circuit design for the algorithm model of low-power voice activity detection including:1.According to the requirements of the classification algorithm,an environmental noise detection algorithm is proposed.By adding the 17th channel of analog feature extraction,the environmental noise detection value is calculated.Based on the judgment of the environmental noise detection value,the working mode of the circuit is selected,including low power mode(BNN used)and high-performance mode(BWN used).2.According to the characteristics of BNN and BWN,their process element(PE)is designed respectively,and the approximation method is used to simplify the calculation of bias,batch normalization and activation circuit,which reduce the calculation power consumption.3.A precision self-adaptive OR gate approximate addition unit is proposed and applied to the BWN calculation unit.While ensuring the calculation accuracy,the power consumption of a single addition unit is reduced by 30%.This thesis is verified with the TSMC 22nm ULL process.The low-power voice activity detection proposed in this thesis owns a recognition rate of 94%at 10dB SNR environment with the power consumption of 0.508?W;a recognition rate of 91.43%at-5dB SNR environment with the power consumption of 0.578?W.In addition,the enhanced voice activity detection proposed in this thesis supports a wake-up word detection,the recognition rate is above 90%at 10dB SNR environment,and the power consumption is 0.881?W.