Large-vocabulary speech recognition using an integrated syntactic and semantic statistical language model

A method for performing speech recognition, comprising the steps of:

• receiving an acoustic signal;
• processing the acoustic signal to form an acoustic feature sequence;
• and
• processing the acoustic feature sequence using a multiple-span stochastic language model to form a linguistic message, wherein the multiple-span stochastic language model includes a local span providing an immediate word context and a large span providing a global word context.

A method for deriving a multiple-span stochastic language model, comprising the steps of:

• reading at least one training document;
• using an n-gram paradigm to compute a set of local a priori probabilities based on said at least one training document;
• using a latent semantic paradigm to compute a set of global a priori probabilities based on said at least one training document;
• and
• combining the local a priori probabilities and the global a priori probabilities to provide the multiple-span stochastic language model.

A computer readable medium containing program instructions for deriving a multiple-span stochastic language model, the program instructions including instructions for:

• reading at least one training document;
• computing a set of local a priori probabilities, based on said at least one training document, using an n-gram paradigm;
• computing a set of global a priori probabilities, based on said at least one training document, using a latent semantic paradigm;
• and
• combining the local a priori probabilities and the global a priori probabilities to provide the multiple-span stochastic language model.