Abstract: 【Objective】In this research, speech recognition technology was applied to collect agricultural price information. The aim of the research is to recognize the continuous speech which is limited in vocabulary and uttered by independent Chinese mandarin speakers, and to propose a robust speech recognition method suitable for the environment where agricultural product prices are collected. On the basis of Hidden Markov Model (HMM), we train the acoustic models for this environment, so as to relieve the decrease of recognition rate caused by the mismatching between the test environment and the training environment, and to make further improvement of the recognition rate. 【Method】 In the stage of acquiring and processing data, we first built the transformation grammar according to certain rules to recognize the limited vocabulary, and this grammar will be used to guide the recording of both train data and test data. Then we select different environments to collect agricultural product prices by different speakers. On this basis, we built a speech corpus in which speech data are artificially segmented with accuracy. In the stage of training model, we choose the continuous mixed density Hidden Markov Model with left-to-rigt and non-jump structure, and extract 39 demension MFCC feature vector from training dataset to train the HMMs. Firstly, we select monophones as the recognition unit to train male HMMs, female HMMs, and male-female mixed HMMs. Taking it into consideration that the monophones are poor in stability and vulnerable to coarticulation, we select context-dependent triphone as the decoding unit to retrain above HMMs. Since the number of triphones models will increase significantly when the triphones are chosen as modeling unit, we use the decision tree clustering to solve the insufficiency of training samples. In the process of building a decision tree, we divide all the finals and initials into different sets by using the phonetic knowledge. For the identification of initials, we appeal to their pronunciation way and place, and for the finals, we resort to their constitution and head vowels. In this way we realize the design of binary value questions. On this basis, we increase Gaussian mixture components to make the model more accurately described. Besides, in order to solve the problem of convolution noise in the communication channel, we adopt the CMN and CVN methods to alleviate the mismatching problem between test environment and training environment. Finally, the male and female HMMs are obtained respectively by training. In the stage of test, for the different models employing different methods mentioned above, we do the test experiments with the same test dataset respectively and obtain the sentence recognition rate, word recognition rate, and accuracy of every different method. 【Result】 The results show that recognition performance of triphone models are superior to monophone models. Both male and female HMMs perform better than the male and female mixed acoustic models. Though decision tree clustering method cannot promote recognition rate significantly, it can reduce the quantity of triphone models evidently. Gaussian mixture components improve the recognition rate on the one hand, but they bring a certain amount of increase in calculation on the other. CMN and CVN methods can significantly improve the performance of identification system. Through the different locations and different speaker test, the methods we have used demonstrated varying degrees improvement in the recognition performance. The ultimate recognition rate was 95.04% for males, and 97.62% for females.【Conclusion】It is feasible to apply speech recognition technology to the collection of agricultural product price information. In this paper, we proposed a method to improve the recognition rate in agricultural product price information acquisition. The experiment results show that the models trained by these methods have a good recognition performance. Furthermore, the approach adopted by our research lays a foundation for the development of the application system in the future.

Key words: speech recognition, agricultural price, information acquisition, CMVN, decision tree clustering