Echolocation: Using Word-Burst Analysis to Rescore Keyword Search Candidates in Low-Resource Languages

<p>State of the art technologies for speech recognition are very accurate for heavily studied languages like English. They perform poorly, though, for languages wherein the recorded archives of speech data available to researchers are relatively scant. In the context of these low-resource languages, the task of keyword search within recorded speech is formidable. We demonstrate a method that generates more accurate keyword search results on low-resource languages by studying a pattern not exploited by the speech recognizer. The word-burst, or burstiness, pattern is the tendency for word utterances to appear together in bursts as conversational topics fluctuate. We give evidence that the burstiness phenomenon exhibits itself across varied languages. Using burstiness features to train a machine-learning algorithm, we are able to assess the likelihood that a hypothesized keyword location is correct and adjust its confidence score accordingly, yielding improvements in the efficacy of keyword search in low-resource languages. </p>

Name thesis:degree_name

Master of Arts

Level thesis:degree_level

Master

Discipline thesis:degree_discipline

Linguistics

Grantor

The Graduate School and University Center of The City University of New York

Year dc:date.available

2014

Author dc:creator

• Richards, Justin

Advisor dc:contributor.advisor

• Andrew Rosenberg

• Computer Sciences
• Linguistics
• Babel
• burstiness
• cache
• keyword search
• spoken term detection
• word-burst

Repository record dc:identifier

https://academicworks.cuny.edu/gc_etds/273

OAI identifier oai:identifier

oai:academicworks.cuny.edu:gc_etds-1272