A kind of improving HMM model and using in the visual speech recognition

Visual speech-lip reading, making the computer understands what do speakers want to express through observing the lip direction of them. The most simply method of lip reading in early stage is to compare between characters from the frozen pictures and templates being stored. It neglects the character is changing with time. This method is very simply, but it only can classify the simple elements not the words, so it couldn't render great serves to speech recognition. Afterwards the adoption of behavioral characteristics is becoming more widespread. Because of the superior of Hidden Markov model (HMM), it can be applied in speech recognition widely. In recent years, it is also used in the research of lip-reading recognition. The classical HMM model makes two hypotheses: hidden expropriation hypothesis: the state at t+1 is only conditioned by the state at t, not the state before; the expropriation hypothesis from hidden state to visible state: the visible state at t only conditioned by the hidden state at t, not the state before. Such kind of hypothesis is not very reasonable in some practical application (such as lip-reading). In some kind condition, the state at t is not only conditioned by t-1, but also t-2. Therefore this thesis revises the assumed condition classical HMM to derive a new HMM model and algorithm, and applying it into lip-reading recognition to increase the discrimination.