19th INTERNATIONAL CONGRESS ON ACOUSTICS
MADRID, 2-7 SEPTEMBER 2007



DIVIDED LISTENING IN AUDITORY DISPLAYS

PACS: 43.71.Rt

Best, Virginia; Ihlefeld, Antje; Mason, Christine; Kidd, Gerald Jr; Shinn-Cunningham, Barbara
Hearing Research Center, Boston University, Boston, MA, USA; ginbest@cns.bu.edu


ABSTRACT
Two experiments examined patterns of performance when listeners were asked to respond to
two spoken messages presented simultaneously. In Experiment 1, the level of one message
was systematically varied relative to the other. In selective listening trials, listeners reported two
keywords from this message. In divided listening trials, listeners were also required to report two
keywords from the other message. Responses to the variable-level message were similar in
selective and divided listening: there was a monotonic influence of the level of the message,
and a beneficial effect of spatial separation of the two sources. Responses to the second
message, however, were relatively unaffected by the level or spatial configuration of the
sources. In Experiment 2, the two messages were equal-level but were systematically degraded
by adding noise. Errors in reporting a particular message were more frequent as the noise level
increased, but this increase in errors was more dramatic for the source reported second in
divided listening trials. Together, these results support the idea that different strategies underlie
the processing of two simultaneous messages. The data are also consistent with the
involvement of a volatile sensory trace in divided listening.

INTRODUCTION
Studies of selective listening to speech show that listeners are generally good at retrieving
information from a talker at a location they are attending, but perform poorly when asked to
recall messages from unattended locations [1]. However, several studies have indicated that
listeners have some capacity to process semantic information from messages outside the
immediate focus of attention (see e.g. [2]) and can perform remarkably well at following two
separated talkers when they are instructed in advance to do so [3].

Broadbent [4] postulated that auditory immediate memory allows listeners to process
simultaneous inputs in a serial fashion. In his model, all incoming sensory information is stored
temporarily in a relatively unprocessed state. Selective attention allows an object to be selected
and processed further (e.g. identification of semantic content). In the case of simultaneous
inputs, it is possible to process one input and then use the sensory trace (if it is still available) to
process the other input. He estimated the sensory trace to last for up to a few seconds. If this
model of divided listening applies to listeners processing simultaneous messages in an auditory
display, there may be differences in how report of the different messages is affected by various
parameters of the display. For example, while spatial cues can greatly enhance selective
listening to one message in a mixture, it is not clear whether spatial cues also influence the
processing of one of the competing messages if it is accessed using immediate memory. In
addition, it is not clear how robustly sounds are represented in the sensory trace. It may be that
the processing of a message via the sensory trace mode may be more susceptible to the quality
of the acoustic input than the processing of a message using selective attention.

Here, two separate experiments are summarised that examined divided listening in auditory
speech displays. They provide some preliminary data about how the report of two messages is
affected by the context of the display. Experiment 1 focused on the effect of spatial separation in
a divided listening task, while Experiment 2 examined whether degradations of the acoustic
stimuli differentially affect performance for the two messages.

19th INTERNATIONAL CONGRESS ON ACOUSTICS – ICA2007MADRID
2
EXPERIMENT 1

Methods
Four listeners (ages 21 to 24) participated in Experiment 1. Stimuli were D/A converted and
amplified using Tucker-Davis System 3 hardware and presented over Sennheiser HD 580
headphones to subjects seated in a sound-treated booth. Subjects indicated their responses
using a graphical user interface.

Speech materials were taken from the Coordinate Response Measure corpus [5], which
consists of sentences of the form “Ready <call sign>, go to <color> <number> now”. Color and
number pairs were always chosen randomly with the constraint that they differed between the
two competing sentences. In order to create a difficult attentional task, the same talker was
used for both sentences. However, to minimize the influence of energetic overlap between the
sentences, they were processed into mutually exclusive frequency bands [6]. The processed
sentences were filtered with head-related transfer functions to simulate sources at a distance of
1m in the horizontal plane in four different spatial configurations: two in which the two talkers
were co-located (at either 0° or 90°) and two in which the talkers were spatially separated (one
at 0° and the other at 90°). One sentence (S2) was presented at the same level (approximately
70 dB SPL) on every trial. The level of the other sentence (S1) was varied relative to S2 by an
amount that was chosen randomly from trial to trial (-40, -30, -20, -10, or 0 dB, as well as +10
dB in the selective task only).

In selective listening trials, listeners were asked to report the color and number keywords from
S1, identified by its specific call sign (‘Baron’). In divided listening trials, the call signs of both S1
and S2 were random and listeners were asked to report the color and number pairs from each
message in any order. In a particular run, listeners either performed the selective or divided
listening task, and the spatial configuration was fixed. For each condition/configuration
combination, 12 runs were completed by each listener. A run consisted of eight repetitions at
each level of S1, for a total of 96 repetitions per data point.



Figure 1. Mean percent correct scores as a function of the level of S1 for A) S1 in
the selective listening task, B) S1 in the divided listening task, and C) S2 in the
divided listening task.


Results
In the selective listening task, a response was scored as correct when the subject reported both
the color and number of S1. Figure 1A shows the across-subject mean percent correct as a
function of the level ratio for each spatial configuration. The error bars show the across-subject
standard error of the means. In all spatial configurations, performance improves as the relative
level of S1 increases. An exception to this arises in the co-located configurations (solid lines),
where performance at 0 dB is actually worse than at -10 dB. This effect has been observed in
previous studies [7,8] and is attributed to increased confusability of the competing sources when
they are equal in level. When the two sources are spatially separated, performance is always
better than for the co-located cases (dashed lines fall above solid lines).