首页>
外国专利>
SPEECH RECOGNITION SYSTEM EMPLOYING DISCRIMINATIVELY TRAINED MODELS
SPEECH RECOGNITION SYSTEM EMPLOYING DISCRIMINATIVELY TRAINED MODELS
展开▼
机译:运用差异化训练模型的语音识别系统
展开▼
页面导航
摘要
著录项
相似文献
摘要
In the speech recognition system disclosed herein, each input utterance isconverted to a sequence of raw vectors. For each raw vector, the systemidentifies that one of a preselected plurality of quantized vectors which bestmatches the raw vector. The raw vector information is, however, retained forsubsequent utilization. Each model of a vocabulary word to be recognized is inturn represented by sequence of states, the states being selected from apreselected group of states. However, for each word module state, there isprovided both a discrete probability distribution function (pdf) and acontinuous pdf characterized by preselected adjustable parameters. A storedtable is provided which contains distance metric values for each combinationof a quantized input vector with model state as characterized by the discretepdfs. Word models are aligned with an input utterance using the respectivediscrete pdfs and initial match scores are generated using the stored table.From well matching word models identified from the initial match scores, aranked scoring of those models is generated using the respective continuouspdfs and the raw vector information. After each utterance, the preselectedcontinuous pdfs parameters are adjusted to increase, by a small proportion,the difference in scoring between the top and next ranking models. Preferably,if a user corrects a prior recognition event by selecting a different wordmodel from the respective selected group, a re-adjustment of the continuouspdfs parameters is accomplished by performing, on the current state of theparameters, an adjustment opposite to that performed with the originalrecognition event and performing on the then current state of the parametersan adjustment equal to that which would have been performed if the newlyidentified different word model had been the best scoring.
展开▼