A training prompts generation algorithm for connected spoken word recognition

Research output: Contribution to conferencePaperpeer-review

Abstract

This paper describes an efficient algorithm to generate compact prompts lists for training utterances. In building a connected speech recognizer such as a connected spoken digit recognizer, we have to acquire speech data with various combinations of the words contexts. However, in many speech databases the lists are made by random generators. We provide an efficient algorithm that can generate compact and complete list of words with various contexts. The algorithm begins with a series of unique digits, which is used for difference series of another digits series of wider context. The process is applied recursively until desired context range is achieved. The algorithm can be generalized to any range of contexts, such as tri-words and four-words. This paper includes proof of optimality and completeness of the algorithm.

Original languageEnglish
Pages1557-1560
Number of pages4
StatePublished - 2002
Event7th International Conference on Spoken Language Processing, ICSLP 2002 - Denver, United States
Duration: 16 Sep 200220 Sep 2002

Conference

Conference7th International Conference on Spoken Language Processing, ICSLP 2002
Country/TerritoryUnited States
CityDenver
Period16/09/0220/09/02

Fingerprint

Dive into the research topics of 'A training prompts generation algorithm for connected spoken word recognition'. Together they form a unique fingerprint.

Cite this