A Spoken Document Retrieval System for TV Broadcast News in Spanish and Basque

Resumen

This paper presents a spoken document retrieval system (Hearch) looking like a conventional search tool, which retrieves audio/video segments based on the automatic transcription of speech contents. The system consists of a back-end that captures, processes and indexes audio/video resources, and a front-end that allows to search contents, configure various modules and display performance statistics through a web interface. An early version of this tool is available (http://gtts.ehu.es/Hearch/), which searches and retrieves segments on TV broadcast news repositories in Spanish and Basque. To evaluate the performance of the system, six manually transcribed TV broadcast news in Spanish and seven in Basque have been used. An
approach based on extending the query with the so called friendly terms has been proposed and evaluated, attempting to minimize the effect of errors introduced by the Automatic Speech Recognition module. This approach led to slight performance improvements.