Active hearing, active speaking

A static view of the world permeates most research in speech and hearing. In this idealised situation, sources don’t move and neither do listeners; the acoustic environment doesn’t change; and speakers speak without any effect of auditory input from their own voice or other speakers. Corpora for speech research and most behavioural tasks have grown to reflect the static viewpoint. Yet it is clear that speech and hearing takes place in a world where none of the static assumptions hold, or at least not for long. The dynamic view complicates traditional signal processing approaches, and renders conventional evaluation processes unrepeatable since the observer’s dynamics influence the signals received at the ears. However, the dynamic viewpoint also provides many opportunities for active processes to exploit. Some of these, such as the use of head movements to resolve front-back confusions, are well-known, while others exist solely as hypotheses. This paper reviews known and potential benefits of active processes in both hearing and speech production, and goes on to describe two recent studies which demonstrate the value of such processes. The first shows how dynamic cues can be used to estimate distance in an acoustic environment. The second demonstrates that the changes in speech production which take place when other speakers are active result in increased glimpsing opportunities at the ear of the interlocutor. INTRODUCTION The listening problem The classical account of the issues faced by listeners has been illustrated by the cocktail party problem (CPP): how do listeners manage to decipher speech in the presence of other sound sources, including competing talkers (Cherry, 1953)? The CPP has inspired both behavioural and computational studies which have focused on the use of cues such as fundamental frequency and interaural time differences, invoking principles such as old-plus-new and continuity to handle the introduction and tracking of new sources. The CPP has led to a focus on the existence of multiple sources and, in algorithmic terms, a welcome move away from the idea, prevalent in speech enhancement, that ‘noise’ is a quasi-stationary interference which can be suppressed. Several corpora based on the CPP now exist, and in a recent evaluation of computational techniques for identifying utterances in the presence of another talker, one approach achieved super-human performance in some conditions (Kristjansson et al., 2006). But is the CPP a reasonable description of the true listening problem? By focusing mainly on the idea of attending to a single source amongst multiple non-stationary sources, the CPP account has downplayed a key aspect of auditory scenes, namely their dynamics. Consider instead the following scenario. You arrive at an airport. Your route takes you through the wide, high check-in hall and then through narrow tunnel-like corridors lined with glass and steel to the more comfortably furnished departure lounge, then down even narrower corridors