Full-sequence training of deep structures for speech recognition

Abstract

A method is disclosed herein that include an act of causing a processor to access a deep-structured model retained in a computer-readable medium, wherein the deep-structured model comprises a plurality of layers with weights assigned thereto, transition probabilities between states, and language model scores. The method can further include the act of jointly substantially optimizing the weights, the transition probabilities, and the language model scores of the deep-structured model using the optimization criterion based on a sequence rather than a set of unrelated frames.

Description

technical field [0001] The invention relates to language recognition technology, in particular to learning technology in deep structured models. Background technique [0002] Speech recognition has been the subject of extensive research and commercial development. For example, speech recognition systems have been incorporated into mobile phones, desktop computers, automobiles, etc. to provide specific responses to speech input provided by users. For example, in a mobile phone equipped with voice recognition technology, a user can speak the name of a contact listed on the mobile phone, and the mobile phone can initiate a call to that contact. [0003] Additionally, many companies are currently using speech recognition technology to assist customers in identifying company employees, identifying problems with products or services, and the like. [0004] Motivated in part by the requirement to exploit some similar properties in human speech generation and perception systems, r...

AI Technical Summary

Problems solved by technology

[0005] Although DBNs generally have higher modeling power than their shallower counterparts, learning in DBNs is difficult, in part because the backpropagation algorithm is often ineffective due to the significantly increased chance of getting stuck in local optima. implement

Images