| 摘要 |
PURPOSE: To make it possible to find a key word that the same speaker has spoken in a previously sound-recorded speech by using a non-key work, i.e., a hidden Markov model(HMM) of a background speech. CONSTITUTION: A background HMM is generated by connecting states in parallel and a typical background model has transition from the 1st blank state on the left side to respective M Gaussian output generation states; and transition probability Pi corresponds to previous probability of a state (i) and self- transition probability di models the continuance of the state (i). Further, a typical key word HMM has self-transition in each state and transition to two following states, and is retrained successively by using the same generation. An HMM network used to find the key word is constituted by connecting the background HMM and key word HMM in parallel and a searching method assumes a key word and a place by a forward search passing through the network and confirms the key word by a backward search passing respective networks individually, thereby finding a start point. |
