Feature compression for distributed speech recognition

Abstract

Over the recent years, the use of Automatic Speech Recognition (ASR) has become common in mobile phones. Due to the limited implementation resources, only very simple ASR applications have been introduced for mobile phones. Even though it is today feasible to implement ASR systems capable of coping with continuous speech input and large vocabularies, speaker-dependent isolated word recognition is still the dominant technology in ASR applications running on mobile phones. The Distributed Speech Recognition (DSR) architecture has been introduced to overcome the implementation related difficulties of ASR systems by distributing the computationally intensive parts of a speech recognizer into the powerful network servers and carrying out only speech parameterization and compression in the terminal side. This thesis investigates the DSR approach to ASR. The particular focus is on the development and implementation of a parameter compression scheme for the DSR architecture. /Kir10