Research On Natural Language Understanding For Instrument Calibration Robot

The speech recognition and natural language understanding is the prerequisite for intelligent control of instrument calibration robot.Because the common commercial speech recognition softwares perform not good enough on the management of instrument calibration instructions,this paper aims at understanding the natural language of calibration robot.From the two aspects of acoustic model and language model,the text analysis the key elements which affect the performance of speech recognition system and deal with the understanding of the instructions.At the same time,considering the expansion of the instructions in application,the speaker adaptation algorithms are introduced to solve the sparsity problem in the process of model training.Firstly,this paper preprocesses the speech signal based on short-time analysis,including pre-emphasis and short-time frame analysis.These processes reduce the noise in speech recording and transmission process,and obtain a large number of short-time quasi steady state frames.A 39 dimensional feature based on MFCC is extracted,including 12 dimensional MFCC coefficient,logarithmic energy value,as well as their first and two order difference coefficients.With this method,the speech signals that cannot be processed directly are transformed to mathematical form.Then,a topological structure of 5 states from left to right with the continuous 5 dimensional mixed Gauss distribution is adopted,based on the acoustic model.Embedded training by the Baum-Welch algorithm,a context dependent Triphone model performs better,based on Monophone model which is independent of the context.Decoded by the token transfer method based on Viterbi algorithm,the accurate rate of the speech recognition is up to 90.2%.Breaking the limitations of purely using acoustic model,fusing the bigram to define the context dependence of words limited by statistical language model,the recognition rate is raised to 98.9%.Assign unique ID numbers to each instruction,implement a simple understanding of the natural language,and translate the voice instruction recognition results into a form that the machine can handle with.Finally,as to the sparseness problem in training data which probably occurs at extensions for instrument calibration in the practical process of the robot,the MAP/MLLR hybrid algorithm is adapted in speaker adaptive calculation,with the adaptive recognition rate of 15.5% up to 85%.As a result,the adaptive design of matching with the feature of new speakers can be realized with a small amount of target speakers.