Research On Speech Control Technology For Mobile Robot

With the development of the artificial intelligence, Robots begin to penetrate everycorner of human life from the original industrial applications, they are becoming more andmore intelligent. However, the interaction with robot has no enough update, in general, mostof interaction method is based on keyboad or button, and it is very inconvenient andinefficient. In order to improve the efficiency of the interaction with robot, some scholarshave begun to explore new ways of interaction, one of the most favorite way is voice. Manytechnology are involved in speech interaction of robot, such as speech recognition, speechsynthesis, semantic understanding and so on, however, this article will focus on speechrecognition.First, this article describes the status of domestic and international technology which isrelevant to this study content, and studies the entire process of the speech recognition. I haveanalyzed the statiscal properties of the speech signal and the noise signal, and found thedifference between the voice signal and the noise signal. I have proposed a new endpointdetection method which based on Higher Order Cumulant, it uses the Third-Order Cumulantand Fourth-Order Cumulant. And then,make experiments in simple background noise andcomplex background noise respectively.Secondly, I make a introduction to common used speech features and conducted adetailed analysis of two parameters which are commonly used in speech recognition, that isthe Mel cepstrum coefficient and the linear prediction cepstrum coefficient. Dynamic timewarping algorithm is a earlier method of speech recognition, The basic dynamic time warpingalgorithm bases on the results of the endpoint detection hardly and requires a great amount ofcalculation. This paper proposes an improved strategy for the two issues, the strategy issegmented dynamic time warping which is benefit to reduce the amount of computation andthe dependence on endpoint detection results.the main idea is to find a known match point andthen use the dynamic programming algorithm calculates from a known point to the endswhich can narrow the search space and reduce the impact of the endpoint detection. Therecognition experiments with ten keywords are completed by using segmented dynamic timewarping algorithm, which show that segmented dynamic time wrapping algorithm in the caseof particular person is better than in the non-specific. Thirdly, I study the advantages and disadvantages of the Continues and Discrete HiddenMarkov model. Continuous Hidden Markov Model calculation is more complex and needmore training data. Discrete hidden Markov model calculation is relatively simple andrequires less training data. However, the feature vectors of the speech is a continuous space.Therefore, this paper presents a method which uses K-MEANS algorithm to quantify speechfeature and then uses discrete Hidden Markov Model. The discrete HMM speech recognitionalgorithm has been achieved and the Keyword HMM models have been created by usingspeech data which is from robot platform, and the HMM models have been tested by usingoff-line data.Finally, the recognition algorithm based on K-MEANS and Discrete HMM has beenachieved on the experimental platfrom. The entire system can be divided into acquisitionmodule, detection module, recognition module and control module. In order to improve thesystem of real-time performance, the multi-threading and message communication methodhas been used. The control method of robot based speech needs to satsify two conditions:1)real-time performance, a speech control command needs to be reaction in a short time;2)Accuracy, it is the basic of speech control system. The experimental platform is a crawlerrobot, with the speed of robot increasing the self-noise is growing. So The speed of robot hasa certain impact on the accuracy of recognition.