The Research Of Speech Recognition Technology Based On FPGA

In recent years,speech recognition technology has been widely used in the mobile device and the PC.Voice appliances have played an important role in our life that we can see them everywhere nearly.However,the voice appliances need speech recognition ASIC to get work.Although,there are some ASIC manufacturers in our nation,but the function of the chips they made is so simple that can not be used in a more high level.There are many methods to build a speech recognition system,the FPGA have excellent performances in computing,cycle of designing and hardware resources.So,it is very suit for complex system designs.In this paper,a speech recognition system that is suit for smart home is designed.After study the formation of speech,the theory of auditory system and the related algorithms,I have designed a module to use for framing windowing.The module consists of four memorys,it can Finish the work of Framing windowing successfully in real time.In order to improve the system of the application value and the experience of users in practical,a endpoint detection system which is well work in door is designed.In the beginning of design,I have analyzed a large number of speech signals,then i designed a endpoint detection method based on short-time energy difference use for describe the Waveform.At last,I completed the design of endpoint detection module by combining the method with short-time energy and short-time zero-crossing rate.In test,the module can eliminate indoor noise interference well and the identification accuracy is as high as 96%.In the second part of the speech recognition system design,the work of FFT is implemented by the IP of Altera company.In the parameter extraction of MFCC module,a group of filters are built and the logic of logarithm has achieved by look-up table.In the simulation test,approximate outputs of the hardware and software had got.In the design of vector quantization module,the codebooks were generate by Matlab because of the complexity of work,the module Implement the function of search only.In the design of speech recognition,algorithm about HMM was learned.According to the feature of HMM,the problem of viterbi underflow and the Initialization of hmm parameter were solved.