Speech keyword spotting system and system and method of creating dictionary for the speech keyword spotting system

Abstract

The application relates to a speech keyword spotting system comprising an input unit used for receiving input speech signals, a feature extracting unit used for extracting at least one feature from the input speech signals, a classifier unit used for classifying the input speech signals into target keywords or out-of-vocabulary based on at least one feature and a dictionary, and an output unit used for outputting the classification result. The dictionary comprises grammar and a lexicon, wherein the grammar comprises target keyword lists and at least one sequence used for two identifiers of the out-of-vocabulary; the lexicon comprises pronunciation representing all candidate keywords of speech subsequence of one or more target keywords in the target keyword lists, and pronunciation representing at least one sequence; and one identifier represents a first class of phone, while the other identifier represents a second class of phone. The first class of phone and the second class of phone are clustered based on inherent features of phones. The first class of phone, the second class of phone and the phones of one or more target keywords are described by models.

Description

technical field [0001] The present invention relates to a phonetic keyword spotting system, and a system and method of creating a dictionary for the phonetic keyword spotting system. Background technique [0002] Speech keyword detection techniques aim to detect and recognize predefined speech keywords in continuous or segmented input speech. Speech keyword detection techniques are used in almost all speech recognition applications, such as call center systems, voicemail categorization, and search by content. [0003] In practical speech keyword detection applications, most parts of the input speech do not include keywords. These parts are called Out-Of-Vocabulary (OOV) segments (ie, non-keywords). If an OOV segment is incorrectly identified as a keyword, this is called a false alarm (FA). A high false positive rate can lead to a poor user experience. [0004] Generally speaking, one of the ways to discard OOV data is to build a proper garbage model for OOV data in addit...

AI Technical Summary

Problems solved by technology

However, since a keyword has at least one garbage model, many garbage models will be used at the time of recognition, which causes heavy computation and storage loads for the phonetic keyword detection system

Another disadvantage is that once a keyword is changed or added, the garbage model used for it should be rebuilt and trained

Images