So far HMM was most successfully and widely used in continuous speech recognition.It assumed the independence of feature vectors for the efficient training and recognition algorithm.However the assumption didn't accord with the actual distribution of speech signals.Alternative models that attempt to over- come this difficulty were proposed.They were usually known by the name segmental models.The dissertation thoroughly disccussed Parametric Trajectory Models(PTM),including tone as segmental features in the model,the search of continuous speech recognition and its application in confidence measures.The main works are as following: We realized the PTM recognition system and analysed its probability expressions. When polynomial rank R is zero PTM degenerated to HMM with explicit duration modeling.In multi-segment modelings,the data fitting experiments verified that PTM has more accurate modeling ability compared with H M M.For background silence we supposed that it had an expected linear trajectory though its signal point was irrelated to time.The duration model was import ant to the recognition of silence with frames 1ess than 5. Tone as segmental feature was applicated in PTM.Its soft integration used pitch as the 14th feature just as MFCC.Emulating its trajectory we got the tone of the speech segment.The math essence of PTM determined that, the soft integration can straightly reflect the distribution characteristic of pitch in space. The hard integration method combined tone model with acoustic model.which utilized the property of PTM as segmental models:its structure allowed segmental feature measurements. PTM had more accurate modeling than HMM at the expense of much higher, computation complexity.To Solve this problem the dissertation proposed Fixedframe Parametric Trajectory Model(FPTM),which re-sampled the points in the normalized-time trajectory to the fixed regions and thus avoided the repeated probability calculations of different time points in different speech segments. FPTM can cut 90 times computation complexity while the digit string accuracy falled 0.5%. PTM was attempted in the work of confidence measures.Two methods Were introduced.one was the application of scores.which can be combined with or substituted for HMM scores.The other was the application of recognition results, which verified HMM result.The former introduced new information and improved the description of speech signal。The latter overcomed the limitation of traditional HMMs that it cannot be compared between sentences.In A search we re-recognized the speech segments to be traced in the word lattice using PTM. HMM result was in the different position in the new recognition sequence and it got different confidence weight.So the priority ranking of tracing paths was altered and the recognition accuracy was improved.In hypothesis testing,on the basis of PTM verification Fisher classifier
修改评论