Informational Masking in Speech Recognition

Abstract

Solving the “cocktail party problem” depends on segregating, selecting, and comprehending the message of one specific talker among competing talkers. This chapter reviews the history of study of speech-on-speech (SOS) masking, highlighting the major ideas influencing the development of theories that have been proposed to account for SOS masking. Much of the early work focused on the role of spectrotemporal overlap of sounds, and the concomitant competition for representation in the auditory nervous system, as the primary cause of masking (termed energetic masking). However, there were some early indications—confirmed and extended in later studies—of the critical role played by central factors such as attention, memory, and linguistic processing. The difficulties related to these factors are grouped together and referred to as informational masking. The influence of methodological issues—in particular the need for a means of designating the target source in SOS masking experiments—is emphasized as contributing to the discrepancies in the findings and conclusions that frequent the history of study of this topic. Although the modeling of informational masking for the case of SOS masking has yet to be developed to any great extent, a long history of modeling binaural release from energetic masking has led to the application/adaptation of binaural models to the cocktail party problem. These models can predict some, but not all, of the factors that contribute to solving this problem. Some of these models, and their inherent limitations, are reviewed briefly here.