EXIT Charts for Turbo Automatic Speech Recognition: A Case Study

Inhalt:
Extrinsic information transfer (EXIT) charts, well known in digital communications, have shown to be a powerful tool for analysis and optimization of turbo channel codes, due to their ability to predict the iterative decoding behavior. Considering the parallels between digital communications and automatic speech recognition (ASR), in this paper we rephrase the EXIT analysis to be applicable also for turbo ASR. Our contribution is threefold: First we transfer the concept of mutual information (based on bit loglikelihood ratio values) to turbo ASR (referring to many states). Second we introduce a modeling approach to create the EXITessential transfer characteristics. Third we present an EXIT chart for a case study, based on a magnitude and phase speech information fusion, showing that EXIT analysis indeed shows potential to analyze turbo ASR schemes.