Research Of Deep Learning Based Command Recognition Algorithm And Its Implementation On DSP Platform

As a convenient human-computer interaction technology,speech recognition has been widely used in smart terminal devices.However,nowadays various kinds of speech recognition technologies with high recognition accuracy are based on deep learning algorithms and have very high computational complexity,so they are basically implemented by cloud servers or GPU servers.In order to improve the applicability and get rid of the dependence of network,especially for delay-sensitive applications,it's valuable and necessary to develop the speech recognition technology based on deep learning that can be implemented on embedded processors.The optimization design and research of command recognition algorithm that is suitable for running on digital signal processor(DSP)are carried out in this thesis firstly.A hybrid recognition model named R-SRU in this thesis is proposed.The fore-stage of the hybrid model uses Long Short-Term Memory(LSTM)or Gated Recurrent Unit(GRU)to extract high-order speech features,and its post-stage uses multi-layer Simple Recurrent Unit(SRU)to realize command recognition.Therefore,on the premise of ensuring the recognition performance,the parallelism of the model is improved and the number of model parameters is reduced.The performance of the proposed hybrid model is evaluated on the command data set provided by Google.Under pure speech environment,the recognition accuracy of the R-SRU hybrid model reaches 95.275%,increased by 0.1%compared with the LSTM model,while the model parameters and multiplication operations are decreased by 37.2%and 37.5%respectively.Under the environment of vehicular noise,restaurant noise and pink noise with SNR of 5dB,R-SRU model can also provide a recognition accuracy of more than 90.8%.An embedded hardware platform based on TI C6655 DSP is designed to implement the proposed R-SRU command recognition model in this thesis.Then,taking G-SRU(R-SRU model with GRU as the fore-stage)as an example,we complete the code clipping and transplantation of G-SRU model on C6655 platform.A series of optimization operations are carried out progressively on the transplanted software,including system memory access optimization,code structure optimization and system-level code optimization,which greatly reduces the computational complexity of the algorithm.The test results show that the optimized G-SRU model for C6655 platform requires 37920 bytes of program memory and 915148 bytes of stack and heap memory.157MCycles are needed to complete the recognition of a single command word,which has the recognition speed of 0.125s under 1.25GHz DSP.