Paper Info
Title
Low-Cost Implementation Of Open Set Keyword Spotting
Abstract
This paper investigates low computational and memory cost approaches to open set keyword spotting for audio document retrieval. In many applications the document vocabulary will be fluid and user specific. An open, rather than pre-defined, keyword set is needed to retrieve these documents. Defining the keyword at the time of the search request raises two issues not found in a fixed keyword system, (i) how to model the new keyword and (ii) how to perform the word spotting pass to achieve a resonable retrieval response time. This paper presents Viterbi-based approaches to both of these. The keyword is modelled by one or more phonetic strings. A phone-level recognizer is used to determine the keyword phone string (KPS) from one or more spoken keyword examples. Generating multiple pronunciations using an nbest recognizer was found to give better word-spotting performance than using the 1-best KPS. A more robust KPS was created from 2 compared to 1 keyword samples. Three techniques are presented to reduce the time required to perform the word-spotting search: approximation of the full keyword plus filler recognition pass using the pre-computed Viterbi filler hypothesis; restricting the search space by dynamically matching the KPS against the filler path; and Gaussian Selection. An overall speed-up of x50 in retrieval time was achieved by combining these techniques. (C) 1999 Academic Press.
Year
DOI
Venue
1999
10.1006/csla.1999.0122
COMPUTER SPEECH AND LANGUAGE
Keywords
Field
DocType
document retrieval,search space
Keyword density,Computer science,Computational linguistics,Keyword spotting,Speech recognition,Artificial intelligence,Natural language processing,Document retrieval,Vocabulary,Spotting,Viterbi algorithm,Open set
Journal
Volume
Issue
ISSN
13
3
0885-2308
Citations 
PageRank 
References 
3
0.42
4
Authors
2
Name
Order
Citations
PageRank
Kate Knill124928.02
S.J. Young2368.09