Research On Spoken Digital Recognition Of A Reservoir Computing System Based On Mutually Coupled VCSELs

With the rapid development of information technology,people’s lifestyles are undergoing tremendous changes,and the way people processing information is continuously updated and improved.Artificial neural networks（ANNs）,which simulate human brain behavior and thinking mode,show good intelligence when dealing with complex tasks such as face recognition,chaotic time series prediction,and speech recognition.As a new type of ANN,Reservoir computing（RC）avoids the training difficulties of traditional ANNs.RC can be divided into spatially distributed RC and delay-based RC according to different structure.Compared with spatially distributed RC,delay-based RC has the advantages of simple structure,low cost and easy implementation,so it receives extra attention.The delay-based RC is composed of a single nonlinear node and a feedback loop in which a large number of virtual nodes are distributed.Semiconductor lasers（SLs）can be used as nonlinear nodes of delay-based RC because of their high relaxation oscillation frequency and abundant nonlinear dynamic behavior under external disturbance.In particular,the delay-based RC system using Vertical-Cavity Surface-Emitting Lasers（VCSELs）as the nonlinear node has advantages of small size,low energy consumption and easy integration.Moreover,its rich nonlinear dynamic characteristics is helpful to achieve low error rate when dealing with complex tasks.In this thesis,a delay-based RC system is proposed based on two mutually delay-coupled VCSELs（MDC-VCSELs）,and the system is used to process the spoken digit recognition task.Because the number of virtual nodes in the coupling loop of RC system based on MDC-VCSELs is more than that in the RC system based on a single VCSEL,the processing rate of the system is improved while the error rate is relatively low.In such a scheme,each spoken digit is preprocessed by two different masks to form two masked matrices,whose subsequent column vectors are connected to the preceding one to form two time-dependent series.Then,they are injected into the main polarization mode of two VCSELs to achieve nonlinear mapping,respectively.The transient states of two VCSELs are sampled as virtual node states for training and testing.The results show that when adjusting the injection intensity k_inj and the coupling intensity k_d to operate the laser at the edge of injection locking state,the performance of the system for spoken digit recognition is better.In addition,the frequency detuning between the injected light and MDC-VCSELs has a great effect on the system performance,but parameter mismatch（linear dichroism and linear birefringence）of two VCSELs in the range of 10%has little effect on the system performance.Under optimized parameters,the system can achieve a word error rate（WER）of 0.02%with a recognition speed of 1.1×10⁷ words per second when tested using a dataset consisting of 5000 samples.