Speech recognition system for mobile devices

From HLT@INESC-ID

João Miranda

Date

  • 15:00, October 3, 2008
  • Room 336

Speaker

Abstract

This presentation will cover the development of a speech recognition system, based in the Audimus ASR system. The three main components of the system (feature extraction, acoustic model, and decoder) were sequentially ported to the target device, creating a number of intermediate Distributed Systems, before the final Embedded System. To achieve this, a number of problems related to the limited processing capabilities of the device had to be overcome. The resulting embedded system presents a word error rate of 7.77% running at 0.71 xRT, in a 13161-word radiology task.