Model constrution in Speech recognition on time and space sampling point of view

Speech time series are manifolds in high dimensional feature space. The models of Speech recognition are to reflect the characteristics of the feature space distribution of time series manifolds. Each state in Hidden Markov Model (HMM) corresponds to an area in feature space, and the number of states in model corresponds to time sampling accuracy of manifolds in a higher level. From time and space sampling point of view, this paper analyses the influences of different states number in HMM, indicating that the proper amount of states can increase the performance. Due to different speak rates, low speak rate brings low information rate and high redundancy, this paper resamples feature space vectors according to different inter frame distance, and updates corresponding time information in transcription simultaneously, experiments show performance of HMM models based on feature space resampling increases when reducing samples at proper inter frame distance.