UEP-driven extended feature extraction for ASR over 3G speech channels

In this paper we present an extended feature extraction procedure for Automatic Speech Recognition (ASR) over 3G UMTS channels [2], within a bitstream-based Network Speech Recognition (NSR) architecture. This procedure takes advantage of the Unequal Error Protection (UEP) policy that is applied by the channel coder that highly protects selected parts of the transmitted information, leaving some parameters almost unprotected. Given that wireless cellular channels are very prone to errors, taking this fact into account, the new front-end produces important improvements in the performance of recognizers, increasing the benefits as the quality of channel worsens.