Research On Auditory Evoked Brain Computer Interface For Speech Recognition

The brain’s electrophysiological activity signal acquisition area is generally located in the cranial cavity of the brain or the brain scalp.Poor acquisition accuracy and low signal recognition rate are the main challenges faced by BrainComputer Interface(BCI)research.Currently,BCI signal acquisition can be divided into invasive and non-invasive,but invasive BCI surgery has a higher risk,and it is easy to cause immune response and callus after the operation.Noninvasive BCI has become a research hotspot of BCI technology due to its noninvasiveness,good time resolution,portability,and low cost owing to its use of scalp electroencephalogram(EEG).In non-invasive BCI,motor imagery BCI requires a lot of training on the subjects themselves and EEG data to achieve the desired value;the character spelling speed of visually evoked BCI is much lower than the speed of natural language communication,and epilepsy may be induced by low to medium frequency stimulation over a long period of time.Auditory evoked BCI can effectively improve the spelling speed of characters due to its features such as no training,stable EEG signal,simple and efficient information transmission efficiency,and strong fatigue resistance,thereby improving the acquisition accuracy and signal recognition accuracy.Aiming at the international academic frontiers and the needs of the social market,in this thesis,the model construction,experimental paradigm of auditory-induced BCI,EEG recognition algorithm,etc.have been studied in depth.The main research results obtained are as follows:(1)Aiming at the problem that subjects are prone to fatigue in traditional BCI,combined with NeuroSky’s TGAM chip to quantify subjects’ concentration into eSense parameters,a hearing-induced BCI model based on the supervision paradigm is proposed,and a hearing-induced BCI hardware platform is built,according to the changes in the eSense value of subjects’ concentration during the experiment and the percentage of discarded data,the accuracy of EEG collection has increased by about 30%.(2)Aiming at the problem of previous data sets that do not consider the degree of concentration,an EEG database based on auditory-induced BCI is established,which includes 10 types of digital speech,20 types of short sentences,and 20 types of long sentences,totaling 50 types of tags.The eSense parameter value is obtained by amplifying the EEG signal,filtering out noise and artifacts,and then using the eSenseTM patented algorithm to calculate it.(3)According to the word or sentence,the auditory evoked BCI for digital speech recognition and the auditory evoked BCI for sentence audio recognition has been designed respectively.For digital speech recognition,aiming at the problems of many glitches,and large jitters in the model training of the long and short-term memory neural network(LSTM),the LSTM-t Dense algorithm is proposed in combination with the Dense layer of 3 different hidden units,and finally mapped to the softmax layer to output the classification results,breaking the problem of many glitches and large jitter,compared with the well-performing Gated Recurrent Unit Network(GRU),the recognition rate has been improved by5%;for sentence audio recognition,in response to the over-fitting problem of the algorithm model,a penalty term(L2 regularization)is added to the loss function of the LSTM output weight,and a three-layer Dense based on penalty long and short-term memory neural network(PLSTM-t Dense)is proposed,the model recognition rate reaches more than 94%,which is a 20% improvement compared with LSTM,effectively solving over-fitting.