Fully Flexible Electromagnetic Vibration Sensor For Bone-Conduction Speech Collection And The Corresponding Signal Enhancement Algorithms

Sensing of mechanical vibration based on flexible electromagnetic sensors is challenging due to the complexity of obtaining flexible magnetic membranes with confined and enhanced magnetic fields.A fully flexible electromagnetic vibration sensor is developed to conduct wearable collection of mechanical vibration with excellent adaptability to complex surface morphology through a suspended flexible magnet enclosed within a multi-layer flexible coil and two annular origami magnetic membranes.The annular membranes not only regulate the overall distribution of the magnetic field and enhance the overall magnetism by 291%,but also greatly increase the range of the magnetic field to cover the entire region of the coil.The sensor offers a broad frequency response ranging from 1 Hz to 10 k Hz and a sensitivity of 0.59 m V/μm at 1.7 k Hz.The fully flexible format of the sensor enables various applications demonstrated by biophysical sensing,motion detection,machine diagnostics,and speech collection.Our wearable flexible sensors can be attached on the neck to measure the vibration of vocal cord during human speaking.However,high-frequency attenuation caused by frequency response of the sensor as well as skin absorption of high frequency sound wave impedes the practical application of the sensor that capture speech based on bone conduction.In this paper,a practical speech enhancement system for flexible sensor is proposed,including speech enhancement algorithms based on neural network and hardware deployment of the enhancement model.The capability of four kinds of neural networks including DNN,LSTM,BLSTM,and CRNN have been evaluated,and the performance of different algorithms deployed on four kinds of online and offline platforms has also been investigated.Experimental results show that BLSTM performs best in improving speech quality,but worst in hardware deployment.It improves STOI by 0.18 to nearly 0.80,which is equivalent to a good intelligibility level,but it introduces latency as well as a large model.CRNN,which improves STOI to about 0.75,ranks second among the four neural networks.It is also the only model that achieves the real-time processing among all four hardware platforms,demonstrating its huge potential for deployment on mobile platforms.This task is one of the first trials to systematically and specifically develop processing techniques for voice signal captured by flexible sensors.The results indicate possibility to realize a wearable lightweight speech collection system based on flexible vibration sensor and real-time speech enhancement to compensate high-frequency attenuation.