The Optimization Algorithm Research Of Continuous Speech Recognition System Based On HMM

Speech is the most natural and convenient tool of human-machine communication. It canspread fast in dark without direction limitation. It is the tool which can’t be replaced by visual ortactile information such as pictures, characters and buttons. With the appearance and developmentof computer, human is always seeking to communicate with computer and machine by language,namely machine can recognize human’s speech.Speech recognition technology has achieved enormous success at the present. There are manyeffective methods such as HMM. But there also exist some model and language modeloptimization, training efficiency and recognition rate. Concentration on improving the adaptabilityto environments and decreasing the computation of the recognition algorithms, this dissertationpresents some new algorithms and new schemes. The main contents and innovative contributionsof this dissertation are as follows:1.This dissertation reviews the basic theory, development and current research status ofspeech recognition technology and points out the existent problems and shortcomings.2.For decreasing the complexity of the algorithms obtaining speech signal featureparameters, this dissertation develops the variable length average magnitude differencefunction and the variable length auto-correlation method to estimate speech pitch.Simulations validate the new methods.3.Atfer analyzing the process of computing the linear spectrum frequencies (LSFs), thisdissertation gives one new method: first computing the intervals containing LSFs, thenexactly searching LSFs by successive bisections and interpolations. Simulations showthe methods have less computation and easier to be realized in real time compared withother methods.4.The speech endpoint detection is an important part of speech recognition. The traditionalmethods of speech endpoint detection using energy or zero-cross rate etc can’t achieveideal results in noisy circumstance. The practice proved that detection by eyes is betterthan by using auto-detection based on traditional features. So using image processingsubstitute for eyes’ recognition can obtain a new endpoint detection method, the paperrealized this method, and the experiment shows better results.5.Aiming at the shortcoming of HMM method, this dissertation employs Tabu Search algorithm to HMM parameters training. According to short term memories andgeneralized heuristic global searching of Tabu search algorithm, HMM of speechrecognition can approach optimization.