Determining the state of dialogue on language models

摘要

Systems, methods, devices, and other techniques for determining a dialogue state corresponding to speech input and biasing a language model based on the determined dialogue state are described herein. In some implementations, a method includes receiving, at a computing system, audio data indicative of speech input and determining a particular dialogue state corresponding to the speech input from among a plurality of dialogue states. . A set of n-grams can be identified that are associated with particular interactive states corresponding to speech input. The language model is biased by adjusting the probability score that the language model indicates for n-grams in the set of n-grams in response to identifying the set of n-grams associated with the particular dialog state corresponding to speech input. It can be hung. The adjusted language model can be used to transcribe speech input.