Integrated speech recognition and semantic classification

A medium comprising instructions that are readable and executable by a computing system, wherein the instructions configure the computing system to train an integrated automatic speech recognition and semantic classification system configured for recognizing and semantically classifying acoustic signals, comprising configuring the computing system to:

• use a speech lattice to determine language model parameters for matching a set of acoustic signals to a set of word sequences;
• use a maximum entropy model to determine semantic classification model parameters for matching the set of word sequences to a set of semantic classes;
• evaluate a classification decision rule configured to match acoustic signals to corresponding semantic classes, incorporating the language model parameters and the semantic classification model parameters;
• determine a total classification loss function indicative of a rate of error in matching the acoustic signals to word sequences to semantic classes;
• weight the language model parameters in the classification decision rule to account for the total classification loss function;
• iteratively update at least one of the language model parameters and the semantic classification model parameters to reduce the total classification loss function that incorporates errors in semantically classifying acoustic signals due to the language model and the semantic classification model;
• and
• provide the classification decision rule incorporating the at least one of the iteratively updated language model parameters and the semantic classification model parameters, for the integrated automatic speech recognition and semantic classification system.

A medium comprising instructions that are readable and executable by a computing system, wherein the instructions configure the computing system to perform an integrated process of automatic speech recognition and automatic speech semantic classification on speech inputs, comprising configuring the computing system to:

• receive a speech input;
• match the speech input to a semantic class and to a word sequence corresponding to the semantic class by applying a semantic classification tool to the speech input, wherein the semantic classification rule selects for the highest joint association score between a semantic class and a word sequence for the speech input;
• and
• provide the semantic classes matched to the speech input to an application configured to provide user output that is dependent on the semantic classes matched to the speech input;
• wherein the joint association score comprises parameters of a language model and parameters of a semantic classification model that have been iteratively trained to reduce a total loss function indicative of a rate of error in semantically classifying speech inputs due to errors in either the language model or the semantic classification model.