Topic-dependent Language Model Switching for Embedded Speech Recognition

From HLT@INESC-ID

Marcos Santos Pérez
Marcos Santos Pérez
Marcos Santos Pérez received the Master of Research degree in Telecommunication Technologies from University of Malaga (Spain) in 2011. He was awarded a Ph.D. Fellowship from the Junta de Andalucía (Andalusian Regional Covernment, Spain) to assist in the research project named AVATAR from to the department of Electronic Technology at the same university. His research interests include Embodied Conversational Agents with particular interest on their implementation and performance on embedded devices.
Addresses: www mail

Date

  • 15:00, Friday, December 2nd, 2011
  • Room 20

Speaker

  • Marcos Santos Pérez, Universidade de Málaga

Abstract

Nowadays, many of the applications for mobile embedded devices have a voice interface that uses Automatic Speech Recognition (ASR). In the case of multi-domain/topic speech recognition, other researchers propose the use of several recognizers in parallel or in series to improve the WER (Word Error Rate). This kind of approach is not feasible in embedded devices because the recognition time would be prohibitive. Another approach would be the use an external server to perform the recognition at expense of certain limitations (network availability, latency, etc.).

This work focuses on a new approach based on Language Model Switching (LMS). In this case, the topic detection and language switching are performed within a multipass ASR system running on an embedded device.


Note: This seminar will be held in English.