| 摘要 |
A speech recognition system is described that enables highly accurate speech recognition by concentrating on the transitional features of speech, without a large amount of calculation or storage of parameters. The system comprises means for generating spectrum data from input speech in every predetermined time interval; means for quantizing said spectrum data by using a predetermined spectrum prototype set for recognition, and for generating a corresponding recognition spectrum prototype identifier of each of said spectrum data; means for generating spectrum variation data from said input speech in said time interval; means for quantizing said spectrum variation data by using a predetermined spectrum variation prototype set for recognition, and for generating a corresponding recognition spectrum variation prototype identifier of each of said spectrum variation data; means for storing a plurality of established models corresponding to speech of said time interval, and identified by model identifiers relating to the spectrum data and those for the spectrum variation data, each of which models has one or more states, transitions from said states, probabilities of said transitions, output probability to output each of said recognition spectrum prototype identifiers in said states or said transitions, and output probability to output each of said recognition spectrum variation prototype identifiers in said states or said transitions; means for relating units to be recognized to a chain consisting of a plurality of probability models; means for generating a likelihood, in which a predetermined unit to be recognized outputs a stream of said recognition spectrum prototype identifiers and a stream of said recognition spectrum variation prototype identifiers generated from unknown input speech, based on said occurrence probabilities and output probabilities of the probability models related to said unit to be recognized; means for outputting the results of the recognition of said unknown input speech based on said likelihood, and means for outputting a result of recognition of said unknown input speech based on said likelihood, a plurality of said probability models having a common output probability of each recognition spectrum prototype identifier as long as said plurality of probability models have a common spectrum data related model identifier, and a plurality of said probability models having a common output probability of each recognition spectrum variation prototype identifier as long as said plurality of probability models have a common spectrum variation data related model identifier. |
