Experimetal Research On Speech Recognition Technology For Embedded Computing

Speech recognition technology has achieved significant progress with many researchers' enormous efforts in the past tens of years, and some algorithms have been successfully applied. Recently, industrial control, electronic products have developed rapidly, and speech recognition technology has great application potential as one of convenient interactive means of human and machine. Based on Hidden Markov Model (HMM), small-vocabulary speaker-independent isolated-word recognition for embedded systems has becoming one of hotspots of current research. This thesis systemically studies various techniques which are related to small-vocabulary speech recognition system. Then finishes the training and testing algorithms on the Matlab experimental platform, obtains relevant experimental dates for software/hardware co-design to realize the speech recognition. Also the realization problem on FPGA is proposed.Theoretical model is significative to help design the speech recognition system. Firstly, this thesis analyzes the framework of speech signal and Chinese syllable characteristic in different processing levels, to choose the syllable as the fundamental element for the speech recognition rationally. And then detailedly discuss the speech recognition process, Matlab-based software experimental platform, using HMM to realize the isolated-word recognition.HMM has advantage of establishing model for time series, so we can use HMM to establish models for each word. Compared with every word's model, a word's speech can be recognized. It is a robustness model, can ensure the high lightness. Therefore, broadly apply to the speech recognition field. In this thesis, we simulate and comply some sub-programs on the platform of Matlab. These sub-programs, which include pretreatment, endpoint detection, characteristic extraction, vector quantization, HMM templates training and searching, realize the whole course of speech recognition. Chinese digital speech recognition experimentation proves this scheme is feasible. At the same time, on the premise of researching classics methods, this thesis leads dynamic length of windows and fuzzy theory to the speech recognition process. The small-vocabulary isolated-word recognition experiment confirms it.Finally, this thesis builds a software/hardware co-design experiment platform using Matlab,VC++ and FPGA. It discusses the speech recognition algorithm application on FPGA, and ladies emphasis on the flow-process decision, fixed point computing of MFCC, Viterbi algorithm.Combining FPGA structure characteristic, we use adder, multiplier, comparator etc, to build a Viterbi algorithm structure directly. Using modified Viterbi scoring procedure and pre-computed logic, we realize a simple speech template matching based on HMM. And the experimentation proves this scheme is feasible.