Spectral contrast enhancement of speech in noise for listeners with sensorineural
hearing...

RESPONSE TIMES


Abstract --This paper describes a series of experiments evaluating the effects
of digital processing of speech in noise so as to enhance spectral contrast,
using subjects with cochlear hearing loss. The enhancement was carried out on a
frequency scale related to the equivalent rectangular bandwidths (ERBs) of
auditory filters in normally hearing subjects. The aim was to enhance major
spectral prominences without enhancing fine-grain spectral features that would
not be resolved by a normal ear. In experiment 1, the amount of enhancement and
the bandwidth (in ERBs) of the enhancement processing were systematically varied.
Large amounts of enhancement produced decreases in the intelligibility of speech
in noise. Performance for moderate degrees of enhancement was generally similar
to that for the control conditions, possibly because subjects did not have
sufficient experience with the processed speech. In experiment 2, subjects judged
the relative quality and intelligibility of speech in noise processed using a
subset of the conditions of experiment 1. Generally, processing with a moderate
degree of enhancement was preferred over the control condition, for both quality
and intelligibility. Subjects varied in their preferences for high degrees of
enhancement. Experiment 3 used a modified processing algorithm, with a moderate
degree of spectral enhancement, and examined the effects of combining the
enhancement with dynamic range compression. The intelligibility of speech in
noise improved with practice, and, after a small amount of practice, scores for
the condition combining enhancement with a moderate degree of compression were
found to be significantly higher than for the control condition. Experiment 4
used a subset of conditions from experiment 3, but performance was assessed using
a sentence verification test that measured both intelligibility and response
times. Scores on both measures were improved by spectral enhancement, and
improved still more by enhancement combined with compression. The effects were
statistically more robust for the response times. When expressed as equivalent
changes in speech-to-noise ratio, the improvements were about twice as large for
the response times as for the intelligibility scores. The overall effect of
spectral enhancement combined with compression was equivalent to an improvement
of speech-to-noise ratio by 4.2 dB.


Key words: compression, hearing impairment, response times spectral enhancements
speech intelligibility.


INTRODUCTION


People with moderate sensorineural hearing impairment often complain of
difficulty in understanding speech in noise. They can understand speech
reasonably well in one-to-one conversation in a quiet room, but they have great
difficulty when there is background noise or reverberation, or when more than one
person is talking. This difficulty appears to be related to a variety of
abnormalities in the perception of sound (1) and it persists even when the speech
is amplified sufficiently (by a hearing aid) to be well above the threshold for
detection (1,2).


Reduced frequency selectivity is a well-documented abnormality that is associated
with sensorineural hearing loss and which can affect speech perception in noise.
Frequency selectivity refers to the ability of the ear to resolve a complex sound
into its frequency components. This ability is often characterized by describing
the ear as containing a bank of overlapping bandpass filters, known as the
auditory filters (3). The characteristics of these filters for normally hearing
people have been reasonably well established (4,5,6,7). Sensorineural hearing
loss, and particularly cochlear hearing loss, is associated with
broader-than-normal auditory filters, that is, reduced frequency selectivity
(8,9). Several studies have shown that the ability to understand speech in noise
is correlated with measures of auditory filter bandwidth, although the effects of
filter bandwidth are difficult to separate from the effects of a simple loss of
sensitivity to weak sounds, since the two are highly correlated (10,11,12). It
seems likely that impaired frequency selectivity is at least partly responsible
for reduced ability to hear speech in noise, although this causal link has not
been universally accepted (13).


One mechanism by which impaired frequency selectivity could affect speech
perception in noise involves the perception of spectral shape. The recognition of
speech sounds requires a determination of their spectral shapes, especially the
locations of spectral prominences (usually formants). One representation of
spectral shape in the auditory system is called the excitation pattern. The
excitation pattern of a given sound may be defined as the magnitude of the
outputs of the auditory filters in response to that sound as a function of filter
center frequency (4,6). The excitation pattern resembles a smoothed version of
the spectrum. Broader auditory filters produce a more highly smoothed
representation of the spectrum. If spectral features are not sufficiently
prominent, they may be smoothed to such an extent that they become imperceptible.
In one study where degree of spectral contrast was varied, the contrast (decibel
[dB] difference between peaks and valleys in the spectrum) required for vowels to
be identified was shown to be greater for impaired than for normal listeners
(14). Adding a noise background to speech fills in the valleys between the
spectral peaks and thus reduces their prominence, exacerbating the problem of
perceiving them for people with broadened auditory filters.


A second possible effect of reduced frequency selectivity on speech perception in
noise is connected with the temporal patterns at the outputs of individual
auditory filters. The perceived frequency of a given formant and/or the
fundamental frequency of voicing may be partly determined by the time pattern at
the outputs of the auditory filters tuned close to the formant frequency (15,16).
Background noise disturbs this time pattern, which may lead to reduced accuracy
in determining these frequencies. This effect would be greater in a person with
reduced frequency selectivity, since broader filters generally pass more
background noise.


If reduced frequency selectivity impairs speech perception, then enhancement of
spectral contrasts might improve it for the hearing-impaired person. Either of
the two mechanisms outlined above, one based on degradation of spectral shape and
the other on degradation of temporal patterns, provides a rationale for
performing spectral enhancement. If spectral features are smoothed by an impaired
auditory system, then preprocessing the signal to enhance spectral contrasts can
produce an excitation pattern that more nearly resembles the excitation pattern
evoked by an unprocessed signal in a normal auditory system. The impaired
auditory system can be thought of as convolving the spectrum with a smoothing
function, and spectral contrast enhancement can be thought of as a partial
deconvolution process. If temporal patterns are disturbed by the noise passing
through a broadened auditory filter, then enhancing those portions of the
spectrum where the signal-to-noise ratio is highest (the peaks) and suppressing
those portions where it is lowest (the valleys) should minimize this effect.


Several authors have described attempts to improve speech intelligibility for the
hearing impaired by enhancement of spectral features. Boers (17) processed a set
of sentences so as to increase the level differences between peaks and valleys in
the spectrum. Noise was added after the processing, and the effects of the
processing were assessed by measuring the speech-to-noise ratio required for 50
percent of the words to be understood. Overall, the processing reduced
intelligibility, although two impaired listeners did show a slight improvement
with the processed signals. Even if it had systematically improved
intelligibility, this kind of processing would not be feasible with naturally
occurring signals; with these the speech would already be contaminated with
noise, and the processing would have to operate on the speech-plus-noise.


Summerfield, et al. (18), synthesized "whispered" speech sounds, and investigated
the effect of narrowing the bandwidths of the formants (spectral resonances) used
in synthesis. Narrowing these bandwidths led to both sharper spectral peaks and
greater peak-to-valley ratios. However, it had only small effects on speech
intelligibility; identification of consonants at the end of syllables tended to
be slightly better for both normal and impaired listeners when the formant
bandwidths were half their nominal normal values. Speech intelligibility in noise
was not tested.


Simpson, et al. (19), described a method of digital signal processing of speech
in noise so as to increase differences in level between peaks and valleys in the
spectrum. Before spectral enhancement, the spectra were smoothed to eliminate
minor peaks and ripples, using smoothing filters based on the properties of the
auditory filters in normal ears. The enhancement was also done on a frequency
scale related to the frequency resolution of normal ears (4). The enhancement
procedure involved convolving the spectrum with a Difference-of-Gaussians (DOG)
filter. This operation is similar to taking a smoothed second derivative of the
spectrum. The spectral pattern obtained in this way was used to construct a gain
function to enhance the original spectrum. The intelligibility of the speech in
speech-shaped noise was measured using subjects with moderate sensorineural
hearing loss. The results showed small but reasonably consistent improvements in
speech intelligibility for the processed speech. The processing used by Simpson,
et al. ran at about 200 times real time on a reasonably fast laboratory computer
(Masscomp 5400 with floating-point accelerator).


Stone and Moore (20) described a speech-processing system similar to that used by
Simpson, et al., but one that was simpler, and based on analog electronics
running in real time, using a 16-channel band-pass filter bank. Each channel
generated an "activity function" that was proportional to the magnitude of the
signal envelope in that channel, averaged over a short period off time. A
positively weighted activity function from the nth channel was combined with
negatively weighted functions from channels n - 2, n - 1, n + 1, and n + 2,
giving a correction signal used to control the gain of the band-pass signal in
the nth channel. Recombining the band-pass signals resulted in a signal with
enhanced spectral contrast. Two different experiments were described, the first
using the activity function as described, and the second using a nonlinear
transform of the activity function. In both experiments, several different
weighting patterns were used in calculating the correction signal. The
intelligibility of speech in speech-shaped noise processed by the system was
measured for subjects with moderate sensorineural hearing loss. In both
experiments, no improvement in intelligibility was found. However, subjective
ratings of the stimuli used in the second experiment indicated that some subjects
judged the processed stimuli to have both higher quality and higher
intelligibility than unprocessed stimuli.


Bunnell (21) described a method of digital signal processing to enhance spectral
contrasts. Contrasts were enhanced mainly at middle frequencies, leaving high and
low frequencies relatively unaffected. Unlike the processing used by Simpson, et
al. (19), and by Stone and Moore (20), the enhancement was performed on a
spectral envelope that was calculated with a linear frequency scale (using a
cepstral smoothing technique) rather than a scale reflecting auditory frequency
selectivity. Small improvements were found in the identification of stop
consonants presented in quiet to subjects with sloping hearing losses. No
measurements of the intelligibility of speech in noise were reported.


Several other authors have described methods of processing speech in noise aimed
mainly at enhancing speech quality and/or intelligibility for normal listeners or
as preprocessors for speech recognition devices. Lim (22) reviews work done prior
to 1983. Many of the techniques that have been developed result in improvements
of signal-to-noise ratio (SNR) without any improvement in intelligibility, and
many have been plagued by artifacts such as the introduction of spurious sounds
as a result of enhancing random spectral peaks. Cheng and O'Shaughnessy (23)
described a method similar to that used by Simpson, et al. (19), but differing in
several details. They reported an improvement in subjective quality for speech in
white noise, based on informal tests with normal listeners. They used two
alternative algorithms--one for low-noise conditions where the improvement in SNR
was modest but speech quality (naturalness) was retained or enhanced, and the
other for high-noise conditions, where there was a large improvement in SNR but
speech quality was degraded. No formal measurements of speech intelligibility
were made.


Clarkson and Bahgat (24) filtered signals into several contiguous frequency bands
and expanded the envelope in each band, so as to enhance spectral contrast. A
measure of spectral variance was used to control the amount of expansion.
Listening trials with a simplified real time system showed small, but reasonably
consistent, improvements at 0-dB speech-to-noise ratio in a modified rhyme test.


In this paper, we describe a series of experiments aimed at further developing
the technique of Simpson, et al. (19). Experiment 1 was a parametric study using
processing similar to that described by Simpson, et al. The objective was to find
optimum values of two of the parameters used in the processing. The
intelligibility of speech in speech-shaped noise was measured for several
different conditions involving spectral enhancement. Experiment 2 was carried out
using a subset of the conditions from experiment 1, to determine whether the
spectral enhancement produced improvements in subjective judgments of speech
quality and intelligibility. Experiment 3 investigated the effect of combining
spectral enhancement with amplitude compression, with a modified enhancement
algorithm, again using measures of the intelligibility of speech in speech-shaped
noise. Finally, experiment 4 used a subset of the conditions from experiment 3,
but performance was evaluated in a test measuring both speech intelligibility and
response time. Although the experiments were primarily concerned with the
intelligibility and quality of speech in noise, informal listening tests were
carried out using speech in quiet. In all cases, the quality of the processed
speech was judged to be good, by both normal and hearing-impaired listeners.


EXPERIMENT 1


Method of Speech Enhancement


The technique used for spectral enhancement was similar to that described by
Simpson, et al. (19), and involved manipulation of the short-term spectrum of the
speech in noise. Sampled segments of the signal were windowed, smoothed,
spectrally enhanced, and then resynthesized using the overlap-add technique (25).
Each step is described below. The steps are also illustrated in Figure 1.


The speech in noise was low-pass filtered at 4 kHz (Fem EF16, 100 dB/oct slope)
and sampled at a 10-kHz rate with 12-bit resolution using a Masscomp 5400
computer with EF12M analog-to-digital converter. A 12.8-ms segment of the signal
was weighted with a 12.8-ms Hamming window; the segment was padded with 64 zeros
at the start and 64 zeros at the end. A 256-point fast Fourier transform (FFT) of
the windowed segment was calculated, giving 128 magnitude values and 128 phase
values. The phase values were stored and subsequent operations were carried out
only on the magnitude spectrum.


To avoid enhancing spectral details that would be undetectable even for a normal
ear, the magnitude spectrum was transformed to an auditory excitation pattern,
using the convolution procedure described by Moore and Glasberg (4). This
involved calculating the output of an array of simulated auditory filters in
response to the magnitude spectrum. Each side of each auditory filter is modeled
as an intensity-weighting function, assumed to have the form of the
rounded-exponential filter described by Patterson, et al. (26):


W(g) = (1 + pg)exp(- pg), [1]


where g is the normalized distance from the center of the filter (distance from
center frequency divided by center frequency, deltaf[sub c]/f[sub c]) and p is a
parameter determining the slope of the filter skirts. The value of p was assumed
to be the same for the two sides of the filter. The equivalent rectangular
bandwidth (ERB) of this filter is 4f[sub c]/p.


The ERBs of the auditory filters were assumed to increase with increasing center
frequency, as described by Moore and Glasberg (4). As a result of this
calculation, the original 128 magnitude values were replaced with 128 new values,
representing a smoothed version of the original spectrum. The smoothing tended to
remove minor irregularities in the spectrum, but to preserve peaks corresponding
to major spectral prominences in the speech.


An enhancement function was derived from the excitation pattern by a process of
convolution with a DoG function (on an ERB scale). This function is the sum of a
positive Gaussian and a negative Gaussian that has twice the bandwidth of the
positive Gaussian, as described by the following equation:


DoG(deltaf) = (1/2phi)[sup 1/2][exp[ -- (Af/b)[sup 2[/2] -(1/2)exp[--
(deltaf/2b)[sup 2]/2] l, [2]


where Af is the deviation from the center frequency, and b is a parameter
determining the bandwidth of the DoG function. Note that the total area of this
function, summed over positive and negative parts, is zero. In these experiments
three values of b were used, chosen so that the width of the positive lobe
(between the zero-crossing points) was either 0.5, 1.0, or 2.0 times the ERB of
the auditory filter with the same center frequency (4). Thus, the width of the
DoG function increased with increasing center frequency. The three bandwidths
used will be referred to as B.5, B1, and B2.


The DoG function was centered on the frequency of each of the 128 magnitude
values of the excitation pattern in turn. For a given center frequency of the DoG
function, the value of the excitation pattern at each frequency (in linear power
units) was multiplied by the value of the DoGfunction at that same frequency, and
the products obtained in this way were summed. The magnitude value of the
excitation pattern at that center frequency was then replaced by that sum.


The enhancement function derived in this way was then used to modify the
excitation pattern. At center frequencies where the enhancement function was
positive, the excitation pattern was increased in magnitude; at center
frequencies where the enhancement function was negative, the excitation pattern
was decreased in magnitude. This was achieved in the following way. Let the
absolute value of the enhancement function at a particular center frequency be
denoted by abs(ENF) and the corresponding sign (positive or negative) of the
enhancement function be denoted sign(ENF). The value of the enhancement function
was converted to a decibel-like quantity by calculating


G = log[abs(ENF) + 1] x sign(ENF). [3]


The value of abs(ENF) was generally large (in the thousands), but 1 was added to
it to avoid the possibility of taking the logarithm of zero. The value of G was
then scaled by a certain factor, E, and added to the magnitude of the excitation
pattern at that center frequency--the excitation level being expressed in
decibels. The degree of enhancement of the spectrum was determined by the size of
the factor E; values used were 0.3, 0.6, and 0.9, corresponding to small, medium,
and large amounts of enhancement. These degrees of enhancement will be referred
to as E3, E6, and E9, respectively.


The magnitude values from the enhanced excitation pattern, expressed in linear
amplitude units, were then combined with the original phase values, and an
inverse FFT was used to produce a 25.6-ms segment of spectrally enhanced speech
in noise. This process was repeated every 6.4 ms, and the resultant overlapping
segments were summed to give a complete processed waveform.


In summary, the processing had the effect of enhancing spectral contrast in the
magnitude spectrum while preserving the phase spectrum. The processing was
performed with three degrees of enhancement (E3, E6, and E9) and three values for
the width of the DoG function (B.5, B1, and B2), giving nine experimental
conditions in total. The condition E3B1 is similar to that used by Simpson, et
al. (19). In addition, two control conditions were used. In one, the speech in
noise was processed through all stages except those involving enhancement. Thus,
the spectrum was smoothed in the conversion to the excitation pattern, but was
otherwise unaltered; this corresponds to processing with the value of E set to 0.
We refer to this condition as E0. In the second control condition, referred to as
NULL, the speech in noise was passed through all stages except the conversion to
the excitation pattern and the enhancement. The conversion to an excitation
pattern has the effect of putting a high frequency emphasis on the spectrum; this
happens because the ERB of the auditory filter increases with center frequency.
Since the NULL condition did not involve conversion to an excitation pattern, the
high frequency emphasis was obtained in this condition by increasing the power
spectrum at a given frequency by an amount proportional to the ERB of the
auditory filter at that center frequency. The overall level of the
speech-plus-noise was equalized for all conditions.


Figure 2 shows an example of the spectra of stimuli processed using conditions
NULL (top panel), E0 (middle panel), and E3B2 (bottom panel). The signal was a
synthesized neutral vowel presented in speech-shaped noise at a signal-to-noise
ratio of 0 dB. The figure shows the long-term-average spectra of the processed
stimuli, not the spectra of individual frames; the effects of the enhancement
processing were generally more pronounced in the latter. Note how the spectral
level between the formants, especially the second and third formants, is
decreased by the processing.


Stimuli


The stimuli were the first 11 lists from the Adaptive Sentence Lists (ASL) (27)
presented in a continuous background of noise with the same long-term-average
spectrum as the sentences. Sentences were presented at 12-sec intervals, leaving
ample time for subject responses. Most subjects were tested at a speech-to-noise
ratio of 0 dB, both speech and noise levels being specified in terms of
root-mean-square pressures. Subjects 4 and 6, who scored poorly at this
speech-to-noise ratio, were tested using a ratio of +3 dB. The score was the
number of key words identified (out of the 45 in each list). Stimuli were
recorded on digital audio tape (DAT) and presented via a Quad amplifier and
Monitor Audio MA4 loudspeaker.


Subjects


Eleven subjects were tested. All were diagnosed as having bilateral sensorineural
hearing loss, probably of cochlear origin. Their audiograms and other relevant
information are presented in Table 1. Most were experienced hearing aid users.


Experimental Design


A Latin Square design was used. All subjects were tested with the 11 ASL lists
presented in the same, ascending order. Each subject was tested once in each of
the 11 conditions, with the order of conditions counterbalanced across subjects.
Thus, for each subject, a different list was used for each of the 11 conditions,
and for each condition a different list was used for each of the 11 subjects.


Procedure


The subject sat in a sound-attenuating room facing the loudspeaker at a distance
of 1.3 m. Seven of the subjects, those who normally wore hearing aids without any
compression circuity or other "signal processing," listened using their own
hearing aids. Initially, they were asked to adjust the volume controls on their
aids to the setting that they would use for normal conversation. Then, they were
presented with ASL list 12 (i.e., not one of the 11 test lists) processed using
condition NULL and the level was varied until they indicated that it was at their
preferred listening level. Subject 3, who normally wore hearing aids
incorporating compression, and subjects 2, 5, and 8, who did not normally use
their aids, were tested unaided; the level of the stimuli was adjusted to their
preferred listening level. The adjustments were usually completed well before the
list was completed. The remainder of list 12 was used as practice. In a few
cases, list 13, also processed in condition NULL, was used for further practice.


Testing proper then started. Subjects were presented with 11 test lists, with a
brief rest between each list. Subjects were told to repeat back as many words as
they could, and to make a guess when they were not sure. They were told that the
task would be quite difficult, and they were not expected to hear every word.


Results


The scores for each subject for each condition and the mean scores across
subjects are shown in Table 2. The mean scores do not differ greatly across
conditions, but tend to be lower for the conditions involving a high degree of
enhancement (E9). To assess the significance of these effects, the data were
subjected to an analysis of variance (ANOVA) with factor condition, with the data
blocked across list number and subject (28). In this analysis, the proportions
correct were transformed using the expression arcsine (proportion correct). This
transform makes the scores follow a normal distribution more closely. The effect
of condition was significant: F(10,90) = 5.06, p < 0.001. The GENSTAT package
used gave estimates of the standard errors of the differences between the mean
scores for the different conditions. These standard errors were used to assess
the significance of the differences between means (28, p. 81). The mean score for
condition E9B.5 was significantly lower (p < 0.01) than the mean scores for all
other conditions. The score for condition E9B1 was significantly lower than the
scores for conditions NULL (p < 0.01), E0 (p < 0.01), E3B2 (p < 0.02), and E6B1
(p < 0.05). Scores for the other conditions did not differ significantly.


Overall, these results are disappointing. in contrast to the results of Simpson,
et al. (19), the processing did not improve speech intelligibility relative to
the control conditions; and a high degree of processing led to a significant
worsening of intelligibility. The processing condition giving the highest scores
was one involving a moderate degree of enhancement, E3B2. If scores for this
condition are compared with the mean scores for the two control conditions, NULL
and E0, we find that seven subjects performed better with the processing and four
performed more poorly.


There may be several reasons why Simpson, et al., found significant improvements
in speech intelligibility with processed stimuli, whereas we did not. The first
possibility is connected with the fact that Simpson, et al., used only one
processing condition and one control condition. They gave subjects two practice
lists (one control and one enhanced) and then tested subjects using six sentence
lists for each condition. This gave subjects a reasonably large amount of
experience with the processed stimuli. In contrast, each subject in our
experiment listened to each condition only once, using a single sentence list. It
may be that subjects require a more extended practice period to get a benefit
from the processing. It should also be noted that the use of six lists per
condition greatly reduces the inherent variability in the data compared with our
use of a single list.


A second possible factor only became apparent after the main part of the
experiment was completed. We discovered that the enhancement process had an
undesired side effect; it tended to produce a high-frequency deemphasis. The
spectral level of frequencies above 600 Hz tended to be 3-6 dB lower in the
experimental conditions than in the control conditions. This may have offset any
potential improvements in intelligibility produced by the enhancement process.


EXPERIMENT 2


Ratings of Speech Quality and Intelligibility


There have been several reports in the past of processing that improves the
subjective quality of speech in noise without improving intelligibility (22).
Previous work involving judgments of speech quality has mainly used normally
hearing subjects, although Stone and Moore (20) reported such measurements for
hearing-impaired subjects. Processing that improves speech quality without
changing intelligibility may be useful as a means of making listening more
pleasant and less effortful. Hence, we decided to investigate whether our
processing led to any improvements in subjective speech quality.


Two tests were performed where six hearing-impaired subjects (subjects 1, 4, 7,
9, 10, and 11 from experiment 1) made pair-wise subjective comparisons between
sentences in noise processed using conditions E0, E3B1, E3B2, E6B2, and E9B2 of
experiment 1, rating them for sound quality in one set of tests and
intelligibility in the other set. In addition, we used a condition resembling
E3B1, but with the unprocessed signal added back to the processed signal. This
had the effect of slightly reducing the amount of enhancement, but also of
somewhat reducing the audibility of undesired side-effects of the processing,
specifically a slight "gurgling" quality. This condition resembles the processing
used by Simpson, et al. (19), and will be denoted by E3B1 + U.


For all 15 possible pairs of processing conditions, 10 pairs of sentences were
compared. On a given trial the same sentence was presented twice, the sentences
differing only in the way they were processed. Five different sentences were
used, taken from ASL list 1, chosen because performance in experiment 1 was
especially poor for these sentences. They were edited so that each sentence was
approximately centered in 3 sec of its masking noise. There was an interval of
500 ms between the noises for the two sentences in a pair. Following the end of
the noise for the second sentence, there were 5.5 sec of silence during which the
subject indicated which sentence had the higher quality or intelligibility. Five
of the sentence pairs were presented as condition A followed by condition B,
while the other five were presented as condition B followed by condition A. The
order of presentation of the sentence pairs was randomized for both comparison of
processing condition and order of presentation within each individual test. All
editing was done digitally using a Masscomp 5400 computer system. Final stimuli
were recorded on digital audio tape (Sony DTC 1000ES).


In the first test, subjects were asked to indicate which sentence in each pair
had the higher sound quality in terms of pleasantness. In the second test, they
were asked to indicate which sentence in each pair they felt was more
intelligible. In addition to making a forced-choice decision, each subject was
asked to make a confidence rating on each trial, by giving a number indicating
how large the difference appeared to be.


For each processing pair, a distance metric was calculated by adding together the
10 (signed) confidence ratings. For each pair of conditions, AB, the sign was
positive if B was selected and negative if A was selected. An analysis of the
results showed that, for each subject, there was a high correlation between the
number of selections and the distance metric; correlations ranged from 0.69 to
0.99, and were typically over 0.8. This indicates that the measures have a fairly
high degree of reliability and internal consistency.


Results


The results are summarized in Table 3 (quality judgments) and Table 4
(intelligibility judgments). Each cell in each table shows the number of B
selections (out of 10) with the distance score in parentheses. For example, in
the quality judgments of subject 7, condition E3B1 was preferred over condition
E0 nine times out of ten, with a distance metric of 5.0. It should be noted that
the results showed a bias for the second sentence in a pair to be preferred over
the first. This bias was controlled for by our procedure of balancing the order
of processing conditions across pairs, but it probably had the effect of somewhat
reducing the overall differences between pairs of conditions.


Two overall measures of the scores for each condition were also calculated. For
the first, the preferences were summed across all five pairs of conditions
involving a given condition. For example, the summed preference score for
condition E0 was equal to the total number of times that condition was preferred
over the other five conditions; the maximum value for this score is 50. For the
second measure, the signed distance measures for a given condition were summed
for all comparisons involving that condition. The signs were chosen so that a
positive score would indicate an overall preference for that condition. These
scores are also shown in Table 3 and Table 4.


Consider first the quality judgments (Table 3). For subjects 7, 9, 10, and 11,
conditions E3B1 and E3B1 + U were preferred over the control condition, E0. This
is apparent both from the numbers of selections and from the distance scores.
Preferences for the other processing conditions varied more across subjects.
Subject 11 preferred enhanced speech over the control condition for all
conditions involving enhancement. Her overall scores were lowest for the control
condition, and highest for the conditions involving the greatest degree of
enhancement (E6B2 and E9B2). For subjects 7 and 9, both the overall preference
scores and the overall distance scores were highest for conditions E3B1 and E3B1
+ U, and lowest for condition E9B2. For subjects 1 and 4, preferences were
clearly lowest for the condition involving the greatest degree of enhancement
(E9B2). For subject 10, preferences were less clear cut, but there was a
consistent trend for the conditions involving enhancement to be preferred over
the control condition. Overall, the results indicate that the quality of the
enhanced speech in noise was generally preferred over that in the control
condition for moderate degrees of enhancement. As the degree of enhancement was
increased, subject 10 showed little change in preference, subject 11 showed an
increase in preference, and subjects 1, 4, 7, and 9 showed a decrease.


Consider now the intelligibility judgments (Table 4). For subjects 4, 7, and 11,
all conditions involving enhanced speech in noise were judged to give higher
intelligibility than the control condition, as indicated both by the numbers of
selections and by the distance scores. The overall preference and distance scores
were lowest for the control condition. Subjects 1 and 10 did not show clear
preferences for any condition, and their distance scores were all rather low.
Subject 9 did not show clear preferences for conditions E3B1 and E3B1 + U
relative to the control condition, but tended to prefer conditions E0, E3B1, and
E3B1 + U over conditions involving large degrees of enhancement (E6B2 and E9B2).


In summary, the results of experiment 2 showed that, for judgments of both
quality and intelligibility, speech in noise processed using a moderate degree of
enhancement was generally preferred over the control condition. The results for
higher degrees of enhancement varied across subjects. Subject 11 preferred the
highest degree of enhancement both for quality and for intelligibility. For
several other subjects, quality decreased for the highest degree of enhancement.


EXPERIMENT 3


Experiment 3 was similar to experiment 1, in that it involved measures of the
intelligibility of speech in noise for several processing conditions. However,
the processing differed from that used in experiment 1 in several ways. The first
difference was in the way that the enhancement was performed. Instead of the DoG
function, a function based on the difference between two rounded-exponential
functions was used. This is equivalent to calculating two excitation patterns and
taking the difference between them.


A second difference between experiments 1 and 3 was in the way that the
enhancement signal was transformed into a gain function to modify the spectral
shape of the signal. The transformation in experiment 3 was tailored to limit the
maximum gain at any frequency to 20 dB (to avoid excessive increases in sound
level) and was scaled so that, most of the time, the gain value was within
reasonable limits. In addition, the enhancement function was applied to the
original magnitude spectrum, rather than to the (normal) excitation pattern. This
meant that only major spectral features were enhanced, but fine-grain spectral
features were not smoothed in the conversion to an excitation pattern.


Finally, experiment 3 differed from experiment 1 by including conditions using
fast-acting compression. This was done because the enhancement processing had the
effect of expanding the dynamic range of the speech in noise. Potentially, this
expansion could create problems for hearing-impaired subjects, who often have
loudness recruitment and an associated reduction in usable dynamic range. The
expansion of dynamic range produced by the enhancement processing might have
offset the potential advantages to be gained from enhancement of spectral
contrast. The compression used in experiment 3 was intended to compensate for the
dynamic range expansion.


Method of Processing


Many of the stages in the processing were the same as used in experiment 1.
Therefore, only the stages that were different will be described. The magnitude
spectrum of a windowed sample of speech in noise was determined as before. An
enhancement function was calculated by convolution of the power spectrum with the
sum of a positive rounded-exponential function and a negative rounded-exponential
function. For both the positive and negative functions, the ERB varied with
center frequency according to equations described by Moore and Glasberg (4). The
positive function had an ERB that was 0.5 times the "normal" value suggested by
Moore and Glasberg, while the negative function had an ERB that was 2.0 times the
normal value. The factors of 0.5 and 2.0 were chosen on the basis of informal
listening tests. The sum of the two functions had a positive lobe whose width was
approximately 0.67 ERB, which is intermediate between the B.5 and B1 values for
the DoG function used in experiment 1. Each of the rounded exponentials was
scaled, by dividing by its own ERB, so that the area under it was unity; thus,
the area under the sum of the positive and negative rounded-exponentials was
always zero.


The enhancement function will be designated D(f). It was converted to a gain
function according to the following rules:


Gain(f) = 10[sup K0.3D(f)] for D(f) </= 0 [4]


Gain(f) = 10-(10-1)10[sup -K0.3D(f) for D(f) >/= 0 [5]


The resulting gain function was used to modify the original magnitude spectrum of
the sample of speech in noise by multiplying the magnitude value at each
frequency by the value of the gain function at that frequency. The form of
equation [5] was chosen so as to limit the maximum gain at any frequency to a
factor of 10 (20 dB). The value of the constant K was chosen to give a degree of
enhancement comparable to the E3 enhancement of experiment 1. We refer to this
processing condition as ENH. Stimuli for the control condition, E0, were obtained
by processing stimuli in the same way but with the constant K set to zero.
Subsequent to the enhancement processing, the stimuli in condition ENH were
digitally filtered (by adjusting the magnitude spectrum prior to calculating the
inverse FFT) so that the long-term-average spectrum of the processed noise
matched the long-term-average spectrum of the noise in the control condition.


Examples of spectra for stimuli processed using conditions E0 and ENH are given
in Figure 3. The stimulus was the same neutral vowel in noise as used for Figure
1. Note that compared with the processing condition E3B2 of experiment 1 (lower
panel in Figure 1), condition ENH gave rise to sharper spectral peaks associated
with the formants, and a greater spectral valley between the third and fourth
formants. The difference can be attributed to the fact that the enhancement
function used in experiment 1 was applied to the (normal) excitation pattern,
whereas the enhancement function in experiment 3 was applied to the original
spectrum.


Four conditions using compression were also run. The compression was implemented
using an algorithm described by Robinson and Huntington (29). It was based on the
use of a 20-ms sliding rectangular window. The rms value of the waveform within
the window was calculated for each position of the window, and that value was
used to calculate a gain function applied to the waveform sample at the center of
the window. The compression took two forms. The first gave a moderate amount of
compression, used a compression ratio of 2 and a compression threshold 10 dB
below the peak value of the speech plus noise. We refer to this condition as
C10/2. The second used a greater amount of compression, with a compression ratio
of 3 and a compression threshold 15 dB below the peak value of the speech plus
noise. We refer to this condition as C15/3. The compression was applied both
alone and following the enhancement processing. This gave two additional
conditions involving both enhancement and compression, ENHC10/2 and ENHC15/3.


In summary, six conditions were tested: the control condition, E0; a condition
involving enhancement alone, ENH; two conditions involving compression alone,
C10/2 and C15/3; and two conditions involving both enhancement and compression,
ENHC10/2 and ENHC15/3. The overall level of the speech-plus-noise was equalized
for all conditions.


Stimuli


The stimuli were lists 13-18 from the Adaptive Sentence Lists presented in a
background of noise with the same long-term-average spectrum as the sentences.
All subjects were tested at a speech-to-noise ratio of 0 dB, both speech and
noise levels being specified in terms of root-mean-square pressures. Subjects
were tested without using their hearing aids. In order to compensate for the lack
of aids, which usually give a high-frequency emphasis, the off-tape signals were
passed through a spectrum shaping network that rolled off at 12 dB/octave below
200 Hz, was "flat" from 200 to 400 Hz, and rose smoothly to + 2 dB at 600 Hz and
+ 15 dB at 4 kHz. This form of spectral shaping is similar to that commonly used
in commercial hearing aids.


The level of the replayed speech in noise was adjusted for each subject to the
value that they found comfortable for everyday conversation in a domestic
environment. Other aspects of the stimuli were the same as for experiment 1.


Subjects


Six subjects were tested, four of whom had been used in experiment 1. All were
diagnosed as having bilateral sensorineural hearing loss, probably of cochlear
origin. They are subjects 8-13 in Table 1. Most were experienced hearing aid
users.


Experimental Design


A double Latin Square design was used. Each subject was tested once in each of
the six conditions, with the order of testing of conditions counterbalanced
across subjects. This was then repeated but with the order of testing "rotated"
so that the order of conditions for a given subject was different for the two
Latin Squares. In each Latin Square, one ASL list was used for each subject and
each condition.


Procedure


The procedure was essentially the same as for experiment 1. Subjects were given
one practice list which was also used for adjusting the noise level.


Results


The raw scores are given in Table 5, which also shows the mean score across
subjects for each condition for each of the two Latin Squares. Inspection of the
data revealed a trend for performance to be better for the second Latin Square
than for the first (i.e., there was a practice effect). Therefore, an ANOVA was
conducted with factors condition and order of testing (first or second Latin
Square) with the data blocked across subjects and lists (28). As for the data of
experiment 1, the proportions correct were transformed using the expression
arcsine(square root of proportion correct). The analysis revealed a significant
effect of condition, F(5,50) = 5.89, p < 0.001, and order of testing F(1,50) =
12.04, p = 0.001. The interaction of condition and order of testing approached,
but did not reach, significance, F(5,50) = 1.79, p = 0.13.


Considering the mean scores for both Latin Squares, the highest scores were
obtained for conditions ENH and C10/2 and the lowest for the conditions involving
the greatest amounts of compression, C15/3 and ENHC15/3. Post-hoc tests,
conducted as described earlier, showed that the mean scores for conditions ENH
and C10/2 were significantly higher than the mean scores for conditions C15/3 and
ENHC15/3 (p < 0.01 in all cases). The mean score for condition ENHC15/3 was also
significantly lower than the mean score for the control condition, E0 (p < 0.01).
Thus, a large amount of compression has deleterious effects. However, the scores
for conditions E0, C10/2, ENH, and ENHC10/2 did not differ significantly from one
another.


It seems reasonable to consider separately the scores for the second Latin
Square, since there was evidence for improvements with practice. For the second
Latin Square, the highest score overall was obtained for condition ENHC10/2, the
condition involving both enhancement and a moderate degree of compression. The
mean score for this condition (92 percent) was significantly greater than that
for the control condition (82.6 percent) (p (0.05). However, it was not
significantly greater than the mean scores for conditions ENH (88 percent) and
C10/2 (86.3 percent). The results of the second Latin Square for conditions E0,
ENH, and ENHC10/2 are shown separately for each subject in Figure 4. For subject
8, the differences between conditions were limited by a ceiling effect; scores
were close to perfect for all conditions. All of the other subjects scored better
in condition ENHC10/2 than in condition E0.


In summary, the results of experiment 3 indicate that a large amount of
compression, either used alone or in combination with spectral contrast
enhancement, has deleterious effects on the intelligibility of speech in noise.
The results showed clear effects of practice, suggesting that subjects may
require time to get used to novel types of processing. The results for the second
Latin Square (i.e., those obtained after a small amount of practice) indicated
that the condition involving the combination of spectral contrast enhancement and
a moderate amount of compression, ENHC10/2, gave a significantly higher mean
score than the control condition.


Discussion


Although the results of experiment 3 suggest that the intelligibility of speech
in noise may be improved by the enhancement of spectral contrasts, especially
when combined with a moderate amount of compression, the effects were small. The
small size of the effects probably arose partly from the lack of experience of
the subjects with the processed stimuli; the results showed clear evidence of
practice effects. This raises a dilemma. We wished to compare performance on
several conditions, but we also wanted to avoid the possibility of subjects
learning the sentence lists through repeated presentations. This latter
requirement meant that it was not possible to give the subjects extensive
practice on each condition.


A second factor that may have limited the size of the effects is related to the
trade-off between accuracy and time/effort. Our subjects .were effectively given
as much time as they wanted to respond after each sentence had been presented. In
difficult listening conditions, subjects may have devoted more effort and/or more
time to the task of identifying each sentence. This would have resulted in
reduced differences between conditions in comparison to the hypothetical
situation where equal effort and/or time were devoted to all conditions.


It has previously been suggested that traditional speech intelligibility scores
access only one component of disability and benefit, and that further information
may be obtained by investigation of response times to speech stimuli
(30,31,32,33). The response time aspects have previously been interpreted in
terms of ease of listening. It may be argued that some or perhaps all of the
benefits of spectral enhancement may accrue not from improvement of
intelligibility, but rather from advantages to the listener in terms of the
decreased difficulty (i.e., decreased effort required) in identifying the speech
signal due to the sharper distinction of spectral cues. The availability of the
sentence verification test, which yields measures of both speech intelligibility
and response times, enabled this idea to be tested directly. A further advantage
of this test is that, after an initial practice period, there is little evidence
for improvements over time, and the materials themselves cannot be memorized.
This made it possible to gather much more data for each subject and condition
than in the earlier experiments.


EXPERIMENT 4


The Sentence Verification Test


The sentence verification test uses a closed vocabulary to construct four-word
sentences from an overall vocabulary of 32 words. There are four alternatives for
the first word in the sentence (LIZ, LYNNE, LEN, BEN), 12 alternatives for the
second word (SOLD, SHOWED, STOLE, STORED, WORE, STITCHED, DROVE, CRASHED,
CRACKED, CORKED, READ, TORE), 12 alternatives for the third word (FOUR, MORE,
TWO, FEW, TWEED, CLOTH, FAST, SPORTS, GLASS, JAM, ROAD, STREET), and four
alternatives for the fourth word (CAPS, CARS, JARS, MAPS). Of the 144
combinations of the second and third words, there are 82 for which there is at
lead one fourth word which makes the sentence unequivocally silly (nonsense) and
at least one fourth word which makes the sentence unequivocally sensible (e.g.,
BEN SOLD STREET MAPS is sensible, while BEN SOLD STREET JARS is silly). Any
combination of a fourth word with a second word-third word pair that may be
considered equivocal with regard to sense/nonsense is not employed in the test.
The eventual sentences require identification of the second, third, and fourth
words in the sentence before a decision regarding the sense/nonsense of the
sentence may be made.


The 32 words were stored as digitized waveform files which were isolated from
sentences spoken by a single male talker and were concatenated to produce the
desired sentences. During the construction of the test, care was taken to ensure
that the intonation contours of the items, and other aspects, such as duration of
voicing, were similar across items. This was done to remove extraneous cues not
directly associated with the intelligibility of the individual word.


Following presentation of the sentence to the listener, the subject was asked to
indicate whether the sentence was "silly" or "sensible" via a touch sensitive
computer screen, and the response time for that decision (verification time) was
recorded. This verification was followed by the identification component, for
which four potential alternatives for the first word in the sentence, four for
the second, four for the third, and four for the fourth were displayed on the
touch sensitive computer screen. The subject was required to identify the
components of the sentence. The test may be run either adaptively (yielding a
signal-to-noise ratio for criterion performance) or at a fixed signal-to-noise
ratio (yielding a percent correct score for the intelligibility component). The
verification component of the test yields a median response time for all or a
subset of the items for the cognitive decision concerning the sense/nonsense of
the sentences. Evaluations of the within-session and between-session stability of
the test for both normally hearing and hearing-impaired subjects has shown that
there are no significant long-term learning effects associated with repeated
administration of the closed vocabulary.


Processing of the Sentence Verification Test Items


Due to hardware constraints, the sentence verification test was available only in
Glasgow. Hence, the stimuli to be processed were recorded in Glasgow, sent to
Cambridge for processing, and then returned to Glasgow, using digital audio tape
(DAT) as the recording medium. The 32 individual words constituting the
vocabulary for the sentence verification test were each recorded at
signal-to-noise ratios of 0, + 3, + 6, + 9, and + 12 dB, where the signal level
was defined as the mean level of the speech peaks, and the noise level (shaped
noise with the same long-term spectrum as the single male speaker) was defined as
the rms level. A 1,000 Hz sine wave was included at the beginning of the
recording to provide a reference level. These recordings were then sent to
Cambridge and processed as described earlier, using three of the processing
conditions from experiment 3: the control condition, E0; the condition involving
enhancement alone, ENH; and the condition involving both enhancement and a
moderate degree of compression, ENHC10/2. The processed stimuli were subjected to
the high-frequency emphasis described for experiment 3, before being recorded on
DAT tape. Each condition was recorded on a separate tape. The tapes were then
returned to Glasgow.


The 15 sets of the 32 words (three conditions by five signal-to-noise ratios)
were each redigitized using a CED 1401 laboratory interface and stored as
individual waveform files. These waveform files were concatenated during testing
to produce the required sentences.


Test Conditions


Because the processing was done with the noise added to the speech, the test had
to be administered at fixed signal-to-noise ratios. Each subject was tested both
unaided and with the level of the speech-plus-noise adjusted to a comfortable
value. According to the experimental design described below, the required
condition and signal-to-noise ratio was identified and a total of 55 sentences
were delivered to the subject via a Grason-Stadler GSI 16 Audiometer and a
Goodmans B41 loudspeaker in a sound-treated room with the subject seated 2 m from
the loudspeaker at 0 degrees azimuth. The first five of these sentences were not
scored, but were regarded as practice within each individual run. The remaining
50 sentences were used, giving a score out of 200 for the identification
component of the test. For the verification component of the test, only those
sentences that were correctly identified (each of the four constituent words in
the sentence identified correctly) and verified (correctly labeled as being
either silly or sensible) were used. The median of the response times for the
verification process using this subset of sentences was then derived. Thus, each
run of the sentence verification test yielded an identification score out of 200
(here expressed as percent correct) and a response time(verification time) for
the decision regarding the sense/nonsense of the sentence.


Subjects


The five subjects were all established users (at least 12 months) of a single
post-aural BE10 series National Health Service hearing aid. The characteristics
of the subjects are shown in Table 1 (subjects 14-18). They all had broadly
symmetric bilateral sensorineural losses of moderate degree, with greater losses
at high frequencies than at low. All subjects had taken part in earlier
experiments using the sentence verification test and were familiar with its form
and configuration.


Experimental Design


The experiment consisted of five sessions for each subject, usually conducted at
weekly intervals. Each session used seven complete runs of the sentence
verification test as configured above. During each session, data were gathered
for a pair of signal-to-noise ratios for each of the three processing conditions
(E0, ENH, and ENHC10/2). An initial complete run for one of the signal-to-noise
ratio/processing conditions was employed as practice, as previous experience with
the sentence verification test suggested that optimal stability is achieved if
this is done. The signal-to- noise ratios for each session were selected from a
blocked design across subjects. Within each signal-to-noise ratio, the order of
the three conditions was selected randomly. During the course of the five
sessions, each subject was tested twice for each of the signal-to-noise ratios
and each of the processing conditions.


Results


To show the overall form of the results, the mean of the two repetitions for each
subject/signal-to-noise ratio/condition combination was taken and then the scores
for the five subjects were averaged. The results are summarized in Figure 5.
Error bars show 95 percent confidence limits. The figure shows the expected trend
of increasing intelligibility and decreasing response times as the
signal-to-noise ratio increases. For the identification component, there appear
to be modest but consistent advantages at most signal-to-noise ratios for both of
the processed conditions over the control condition (E0). For the response time
component, the advantages of the processing conditions are larger, relative to
the confidence limits, and there is a clear tendency for the processing condition
involving both enhancement and compression (ENHC10/2) to give shorter response
times than the condition involving enhancement alone (ENH).


The results of the five subjects were subjected to a repeated-measures ANOVA,
using the GENSTAT package, with the following dependent variables: (i) percent
correct score; (ii) arcsine(square root of-proportion correct)--this measure
makes the scores follow a normal distribution more closely; (iii) response time;
and, (iv) square root of response time--again, this measure makes the scores
follow a normal distribution more closely.


The results for the transformed variables (ii) and (iv) were similar to those for
the untransformed variables (i) and (iii), so the latter will be presented to
facilitate interpretation. In the ANOVA, there were three within-subject factors.
The independent variables were: (i) the signal-to-noise ratio (0, 3, 6, 9, and 12
dB); (ii) the condition (linear, enhanced, enhanced and compressed); and, (iii)
replication (first and second replicate).


For the percent correct scores, there was a highly significant effect of
signal-to-noise ratio [F(4,16) = 97.1, p < 0.001], as expected from Figure 5, and
a significant effect of condition [F(2,8) = 6.65, p < 0.021. The main effect of
replicate was not significant, and none of the interactions was significant. The
mean score for condition ENH was 1.76 percent greater than that for condition E0
(standard error = 0.54), and this difference was statistically significant (p <
0.02). The mean score for condition ENHC10/2 was 2.73 percent greater than that
for condition E0, and again this difference was significant (p < 0.001). The mean
difference between conditions ENH and ENHC10/2, 0.97 percent, was not
significant.


The ANOVA for the response time component of the sentence verification test
showed highly significant effects of signal-to-noise ratio [F(4,16) = 333.6, p <
0.001] and of condition [F(2,8) = 31.4, p < 0.001]. The main effect of replicate
was not significant, but there was a significant interaction between
signal-to-noise ratio and processing condition [F(8,32) = 4.07, p < 0.002],
consistent with the greater effect of condition at low signal-to-noise ratios
apparent in Figure 5. The mean response time for condition ENH was 62.8 ms less
than that for condition E0 (standard error = 10.1 ms), and this difference was
statistically significant (p < 0.001). The mean response time for condition
ENHC10/2 was 113 ms less than that for condition E0, and again this difference
was significant (p < 0.001). The mean difference between conditions ENH and
ENHC10/2, 52.8 ms, was also significant (p < 0.001). Thus, the results show that
there are significant advantages for the processed conditions compared with the
control condition, with combined enhancement and compression giving bigger
advantages than enhancement alone. The advantages are statistically more robust
for the response-time component of the test than for the identification
component.


The magnitudes of the effects described above, especially the response times, are
difficult to interpret because of the somewhat complex nature of the sentence
verification test. One way of relating the effects to other, more familiar
measures, is to convert the differences in percent correct scores or response
times to equivalent changes in signal-to-masker ratio. The data in Figure 5
indicate that both the percent correct scores and the response times are
approximately linearly related to the signal-to-noise ratio, for signal-to-noise
ratios between 0 and +6 dB. For the control condition, each 1-dB increment in
signal-to-noise ratio produces a 2.3 percent change in the percent correct score
and a 38.3-ms change in the response time. These relationships were used to
transform the magnitudes of the differences between conditions into equivalent
changes in signal-to-noise ratio in dB.


For the percent correct scores, the difference between conditions E0 and ENH was
equivalent to a 0.8-dB change in signal-to-noise ratio, while the difference
between conditions E0 and ENHC10/2 was equivalent to 1.2 dB. For the response
times, the difference between conditions E0 and ENH was equivalent to a 1.6-dB
change in signal-to-noise ratio, while the difference between conditions E0 and
ENHC10/2 was equivalent to 3.0 dB. Thus, the benefits of processing are
approximately twice as large for the response-time component as for the
identification component. If percent correct and response time can be regarded as
subcomponents of an overall benefit from processing, then condition ENH gave an
overall benefit of 2.4 dB compared with the control condition, and condition
ENHC10/2 gave an overall advantage of 4.2 dB compared with the control condition.


The fully factorial, repeated-measures nature of the experimental design enabled
individual differences to be investigated. For the identification scores, a
general linear model (GLIM) analysis was conducted based on a logistic model
assuming that errors were distributed according to a binomial distribution. Here
the proportion correct (PrC) is the dependent variable in an equation of the
form:


PrC = 1/(1 + exp(- (B0 + B1*X1 + B2*X2 +. . .))) [6]


where X1, X2, etc. are indices for specific values of the independent variables
(signal-to-noise ratios, and processing conditions) and their interactions. The
procedure produced estimates of the values of the parameters, B0, B1, B2, etc.,
referenced to a specific baseline, namely the mean for the control condition at 0
dB signal-to-noise ratio. B0 is the parameter estimate for the baseline itself.
The effect of replicate was not significant and is not included in the analysis.
The results are summarized in Table 6.


To return to a percent correct score from a parameter estimate, the equation


Percent correct = 100/(1 + e- (sum of estimates)) [7]


is used. Thus, for subject 14 the figure of 0.315 for the baseline (control
condition at 0 dB signal-to-noise) is equivalent to a score of 57.8 percent.
Using the properties of the logistic regression, the effect of a combination of
factors may be assessed by simple addition of the parameter estimates. Thus the
estimate associated with a signal-to-noise ratio of 3 dB in the control condition
is 0.315 + 0.282 = 0.597 (equivalent to a percent correct score of 64.5 percent)
while the estimate associated with a signal-to-noise ratio of 3 dB in the
enhanced condition for subject 14 is 0.315 + 0.282 + 0.169 = 0.766 (equivalent to
a percent correct score of 68.3 percent).


The data in Table 6 suggest that the processing conditions give different effects
for different subjects. For example, subject 16 showed no benefit for enhancement
alone, but showed a clear benefit for enhancement with compression. In contrast,
subject 15 showed a clear benefit for enhancement alone, and showed less benefit
for enhancement with compression. Subjects 14, 17, and 18 showed some benefit
from enhancement alone, and showed larger benefits from enhancement with
compression, although the differences between conditions ENH and ENHC10/2 were
not significant. The interaction of signal-to-noise ratio with processing
condition was significant only for subject 17.


The results for the response time estimates were analyzed using an identical
linear model but assuming that errors were normally distributed. The results for
the five subjects are shown in Table 7. The pattern was similar to that for Table
6, though now there was a significant interaction between signal-to-noise ratio
and condition for subjects 16 and 18. Overall, the effects of condition were more
robust (as can be seen by comparing the parameter estimates with their associated
standard errors). For subjects 14 and 15, the benefit of condition ENH was not
significant, but the benefit of condition ENHC10/2 was significant. For subjects
16, 17, and 18, there were significant benefits in both conditions. The benefits
tended to be larger in condition ENHC10/2 than in condition ENH, but the
differences were not significant. It is noteworthy that the identification scores
of subject 16 did not show a benefit for condition ENH, whereas the response-time
scores did.


Although the experiment contained relatively small numbers of subjects, the
nonhomogeneous pattern of results does suggest that, in future experiments, it
would be worthwhile to investigate further the characteristics of individual
subjects to try to find the predictors of benefit from enhancement.


GENERAL SUMMARY, DISCUSSION, AND CONCLUSIONS


The results of experiment 1 were disappointing, in that they failed to show any
significant benefits of the enhancement processing, although the results did
indicate that a large degree of spectral enhancement can have deleterious
effects. In hindsight, the failure to find positive effects of the processing in
experiment 1 probably can be attributed in part to the experimental design, which
did not give subjects the opportunity to practice in the different conditions.


The results of experiment 2 showed that subjective ratings of both quality and
intelligibility were affected by the processing, but the effects varied across
subjects. For judgments of both quality and intelligibility, speech in noise
processed using a moderate degree of enhancement was generally preferred over the
control condition. The results for higher degrees of enhancement varied across
subjects. Subject 11 preferred the highest degree of enhancement both for quality
and intelligibility. For several other subjects, quality decreased for the
highest degree of enhancement.


The results of experiment 3 indicated that a large amount of compression, either
used alone or in combination with spectral contrast enhancement, had deleterious
effects on the intelligibility of speech in noise. The results showed clear
effects of practice, suggesting that subjects may require time to get used to
novel types of processing. The results for the second Latin Square (i.e., those
obtained after a small amount of practice) indicated that the condition involving
the combination of spectral contrast enhancement and a moderate amount of
compression, ENHC10/2, gave a significantly higher mean score than the control
condition.


Taken together, the results of experiments l, 2, and 3 indicate that high degrees
of enhancement, or high degrees of compression, generally have deleterious
effects. In other words, too much processing is a bad thing! However, the results
of experiment 2 indicate that a moderate amount of spectral enhancement can lead
to improved subjective ratings of quality and intelligibility, and the results of
experiment 3 indicate that, after some practice, a moderate degree of spectral
enhancement, combined with a moderate degree of compression, can give better
results than those obtained with unprocessed speech.


Experiments 1 and 3 suffered from the problem that subjects were given rather
little practice in each condition. This was forced upon us because, with the
limited number of sentence lasts available to us, it would not have been possible
to give extensive practice without subjects memorizing the lists. The limited
number of sentence lists created a second problem; it was impossible to gather a
large amount of data for each condition. This meant that some of the effects
observed were of marginal statistical significance. A third problem was that the
measure used, the percent correct of words identified in short sentences, may not
have been suitable for revealing all of the effects of the processing.
Specifically, the measure probably did not tap the dimension of "ease of
listening," which can be especially important in everyday situations involving
decision making and selective attention.


The Sentence Verification Test used in experiment 4 was intended to overcome
these problems. The test can be administered repeatedly without substantial
learning effects, and it includes a measure of response time which is probably
related to ease of listening. The results showed highly significant benefits of
the processing, with spectral enhancement alone being superior to the control
condition, and enhancement combined with compression being superior to
enhancement alone. When expressed in terms of equivalent changes in
signal-to-masker ratio, the benefits were about twice as great for the response
time measures as for the identification scores, and they were also statistically
more robust for the response time measures. This suggests that the major benefits
of the processing may be in terms of increased ease of listening rather than in
intelligibility.


The results of experiment 4 indicate that the improvement in the intelligibility
score produced by processing alone was equivalent to a change in signal-to-noise
ratio of about 0.8 dB, a relatively modest amount. The results of Simpson, et al.
(19) for spectral processing with a similar degree of enhancement (although
implemented using a somewhat different algorithm), showed typical improvements in
intelligibility, relative to the control condition of about 7 percent. For the
speech materials used by Simpson, et al., each 1-dB change in speech-to-noise
ratio produces about an 11 percent change in intelligibility (34). Thus, the 7
percent change in intelligibility is equivalent to about a 0.6-dB change in
speech-to-noise ratio. This is comparable to the 0.8-dB change found in
experiment 4.


It should be emphasized that the overall effect of the processing found in
experiment 4 was larger than this. If the changes in intelligibility and in
response times were both expressed in terms of equivalent change in
speech-to-noise ratio, the net effect was an improvement (relative to the control
condition) of 2.4 dB for enhancement alone, and 4.2 dB for enhancement combined
with compression.


Experiments 1 and 3 used a Latin Square design, which makes it difficult to
analyze the effects of individual differences. However, the results of experiment
2 showed clear evidence of individual differences in the judged pleasantness and
intelligibility of the processed stimuli. Similarly, both the intelligibility
measures and the response time measures of experiment 4 revealed clear individual
differences. Further research is needed to clarify why these differences occur,
and to establish whether they can be related to individual differences in
psychoacoustic factors such as frequency selectivity.


ACKNOWLEDGMENTS


This work was supported by the Medical Research Council (UK) and the Hearing
Research Trust (UK). We thank Brian Glasberg for assistance during many stages of
this work and Adrian Davis for assistance with statistical analysis. Joseph
Alcantara gave helpful comments on an earlier version of this paper.


Table 1. Characteristics of the hearing impaired subjects used in the
experiments.


Frequency in kHz Subject Age Sex Ear 0.25 0.5 1.0 2.0 4.0 8.0


1 74 M L 35 30 35 40 55 55 R 30 35
35 50 65 60


2 75 F L 25 10 5 30 60 75 R 15 10
5 35 65 80


3 67 M L 60 55 55 60 65 75 R 50 50
55 60 70 70


4 68 M L 40 65 75 85 85 >100 R 55 60
70 85 95 >100


5 72 M L 30 20 35 55 95 80 R 35 30
40 60 85 95


6 82 M L 85 80 80 90 95 >100 R 60 65
80 85 80 85


7 70 M L 25 30 60 60 75 80 R 10 10
60 55 65 80


8 78 M L 40 50 55 60 80 >100 R 25 25
40 45 70 80


9 69 F L 35 35 45 45 65 85 R 65 50
55 70 90 100


10 68 M L 70 55 45 35 50 60 R 30 35
45 40 40 50


11 73 F L 45 50 50 50 55 55 R 50 50
55 50 55 55


12 71 M L 30 40 55 30 50 95 R 30 40
45 30 30 60


13 62 M L 45 50 65 60 40 70 R 35 40
55 60 60 55


14 68 M L 30 40 45 60 70 85 R 20 35
45 50 55 75


15 72 F L 35 45 50 55 60 70 R 30 25
40 55 65 65


16 63 M L 15 30 40 60 65 55 R 20 25
40 70 70 60


17 64 M L 25 15 20 40 65 70 R 30 20
20 30 60 75


18 69 M L 20 40 30 35 60 80 R 15 30
40 40 55 75




Subjects 1-11 took part in Experiment 1. Subjects 1, 4, 7, 9, 10, and 11 took
part in Experiment 2. Subjects 8-13 took part in Experiment 3. Subjects 14-18
took part in Experiment 4. Absolute thresholds are given in dB HL.


Table 2. Results of Experiment 1, showing the score for each subject in each
condition (number of words correct out of 45) and the mean score for each
condition.


Legend for Chart:


A - Subject B - Null C - Condition, E0 D - Condition, E3B.5 E - Condition, E3B1 F
- Condition, E3B2 G - Condition, E6B.5 H - Condition, E6B1 I - Condition, E6B2 J
- Condition, E9B.5 K - Condition, E9B1 L - Condition, E9B2


A B C D E F G H I J K
L


1 41 41 40 40 33 37 43 34 30 38
33


2 38 42 38 40 42 37 42 44 30 36
38


3 36 30 35 32 36 30 31 31 16 22
38


4 34 36 31 23 34 33 34 28 19 28
31


5 37 42 41 39 31 30 37 37 19 38
31


6 34 34 33 29 35 33 27 23 22 26
24


7 36 32 36 38 35 28 34 38 35 36
41


8 45 40 34 40 43 35 41 37 23 26
24


9 30 36 31 40 32 38 35 29 31 37
36


10 28 34 29 28 32 32 35 35 36 34
33


11 40 34 40 35 44 36 33 39 35 33
38


Mean 36.3 36.5 35.3 34.9 36.1 33.6 35.6 34.1 26.9
32.2 33.4




Table 3. Results of Experiment 2 for the judgments of the quality of speech in
noise.


Legend for Chart:


A - Subject B - A condition C - B condition, E3B1 D - B condition, ENH E - B
condition, E3B2 F - B condition, E6B2 G - B condition, E9B2 H - B condition,
Global


A B C D E F G
H


1 E0 1 (-4) 5 (1) 3 (-2) 2 (-6)
0 (-11) 39 (22)


E3B1 -- 8 (1) 3 (-1) 1 (-5) 2 (-7)
27 (8)


ENH -- -- 4 (-1) 0 (-5) 1 (-10)
38 (18)


E3B2 -- -- -- 3 (-4) 2 (-8)
25 (8)


E6B2 -- -- -- -- 0 (-8)
16 (-12)


E9B2 -- -- -- -- --
5 (-44)


4 E0 4 (1) 7 (9) 6 (5) 3 (-7)
1 (-19) 29 (11)


E3B1 -- 3 (0) 5 (-2) 4 (2) 1 (-16)
31 (17)


ENH -- -- 6 (2) 7 (5) 1 (-16)
26 (18)


E3B2 -- -- -- 5 (-4) 2 (-8)
30 (17)


E6B2 -- -- -- -- 3 (-5)
26 (50)


E9B2 -- -- -- -- --
8 (-64)


7 E0 9 (5.0) 6 (2.5) 4 (-0.5) 2 (-4.0)
1 (-6.0) 28 (3)


E3B1 -- 5 (-0.5) 4 (-2.5) 2 (-3.5) 4
(-1.5) 34 (13)


ENH -- -- 4 (-1.5) 4 (-2.5) 3 (-4.5)
30 (10.5)


E3B2 -- -- -- 2 (-5.) 3 (-3.0)
27 (3.5)


E6B2 -- -- -- -- 4 (-2.5)
15 (-12.5)


E9B2 -- -- -- -- --
15 (-17.5)


9 E0 9 (17) 8 (13) 3 (-8) 3 (-8)
0 (-26) 27 (12)


E3B1 -- 4 (-2) 6 (4) 4 (-7) 0 (-28)
35 (50)


ENH -- -- 2 (-11) 3 (-10) 0 (-25)
37 (57)


E3B2 -- -- -- 4 (-6) 1 (-18)
26 (9)


E6B2 -- -- -- -- 0 (-21)
24 (-10)


E9B2 -- -- -- -- --
1 (-118)


10 E0 7 (4) 8 (4) 7 (3) 9 (7)
6 (3) 13 (-21)


E3B1 -- 4 (-1) 6 (2) 4 (2) 5 (-2)
28 (3)


ENH -- -- 4 (-2) 5 (0) 5 (-1)
28 (6)


E3B2 -- -- -- 6 (-1) 3 (-1)
28(5)


E6B2 -- -- -- -- 5 (0)
29 (8)


E9B2 -- -- -- -- --
24 (-1)


11 E0 9 (13) 10 (18) 10 (21) 10 (18)
9 (16) 2 (-86)


E3B1 -- 6 (2) 5 (-3) 8 (11) 8 (12)
22 (-9)


ENH -- -- 5 (2) 7 (7) 10 (19)
24 (-8)


E3B2 -- -- -- 9 (11) 9 (14)
22 (-5)


E6B2 -- -- -- -- 6 (2)
38 (45)


E9B2 -- -- -- -- --
42 (63)


Mean E0 65 73 55 48
28 46


E3B1 -- 50 48 38 33
59


ENH -- -- 42 43 33
61


E3B2 -- -- -- 48 33
53


E6B2 -- -- -- -- 30
50


E9B2 -- -- -- -- --
32




Table 4. Results of Experiment 2 for the judgments of intelligibility.


Legend for Chart:


A - Subject B - A condition C - B condition, E3B1 D - B condition, ENH E - B
condition, E3B2 F - B condition, E6B2 G - B condition, E9B2 H - B condition,
Global


A B C D E F G
H


1 E0 6 (1) 4 (0) 7 (2) 5 (0)
4 (0) 24 (-3)


E3B1 -- 3 (-3) 6 (0) 3 (-1) 5 (1)
29 (4)


ENH -- -- 7 (0) 2 (0) 6 (1)
22 (-4)


E3B2 -- -- -- 4 (-1) 5 (0)
31 (3)


E6B2 -- -- -- -- 7 (2)
17 (-4)


E9B2 -- -- -- -- --
27 (4)


4 E0 9 (8) 6 (0) 8 (11) 8 (11)
9 (16) 10 (-46)


E3B1 -- 5 (-2) 5 (0) 5 (7) 4 (-3)
30 (6)


ENH -- -- 4 (0) 5 (5) 5 (1)
27 (-8)


E3B2 -- -- -- 6 (1) 6 (3)
25 (7)


E6B2 -- -- -- -- 4 (1)
30 (23)


E9B2 -- -- -- -- --
28 (18)


7 E0 7 (2.5) 7 (5.0) 9 (6.5) 9 (5.5)
7 (4.5) 11 (-24)


E3B1 -- 2 (-1.5) 4 (0.5) 6 (1.0) 7 (3.5)
28 (-1)


ENH -- -- 7 (2.0) 7 (1.0) 6 (0.5)
19 (0)


E3B2 -- -- -- 7 (3.5) 6 (2.5)
27 (3)


E6B2 -- -- -- -- 8 (2.0)
31 (9)


E9B2 -- -- -- -- --
34 (13)


9 E0 4(-9) 5 (2) 4(-6) 3 (-11)
1 (-18) 33 (42)


E3B1 -- 6 (3) 6 (0) 4 (-7) 1 (-19)
27 (14)


ENH -- -- 3 (-11) 4 (-6) 1 (-24)
33 (46)


E3B2 -- -- -- 4 (-4) 2 (-21)
27 (8)


E6B2 -- -- -- -- 4 (-8)
21 (-20)


E9B2 -- -- -- -- --
9 (-90)


10 E0 6 (1) 4 (0) 7 (2) 5 (0)
4 (0) 24 (-3)


E3B1 -- 3 (-3) 6 (0) 3 (-1) 5 (1)
29 (4)


ENH -- -- 7 (0) 2 (0) 6 (1)
22 (-4)


E3B2 -- -- -- 4 (-1) 5 (0)
31 (3)


E6B2 -- -- -- -- 7 (2)
17 (-4)


E9B2 -- -- -- -- --
27 (4)


11 E0 10 (13) 10 (11) 10 (15) 10 (17)
10 (20) 0 (-76)


E3B1 -- 6 (0) 5 (0) 8 (7) 9 (13)
22 (-7)


ENH -- -- 7 (2) 9 (8) 8 (9)
22 (-8)


E3B2 -- -- -- 10 (9) 9 (14)
23 (-6)


E6B2 -- -- -- -- 7 (2)
40 (39)


E9B2 -- -- -- -- --
43 (58)


Mean E0 70 60 75 67
58 34


E3B1 -- 42 53 48 52
55


ENH -- -- 58 48 53
48


E3B2 -- -- -- 58 55
55


E6B2 -- -- -- -- 62
52


E9B2 -- -- -- -- --
56




Table 5. Results of Experiment 3, showing the score for each subject in each
condition (number of words correct out of 45) for the first and second tests, and
the mean score for each condition.


Legend for Chart:


A - Subject B - Test C - E0 D - C10/2 E - Condition (C15/3) F - Condition (ENH) G
- ENHC10/2 H - ENHC15/3


A B C D E F G H


8 1 40 42 37 37 36 34 2 44
45 42 43 44 38


9 1 41 43 32 41 35 32 2 36
43 36 37 42 33


10 1 32 30 39 42 30 23 2 35
42 43 41 41 27


11 1 41 40 30 38 41 31 2 37
43 38 41 44 38


12 1 42 43 40 41 33 41 2 42
36 36 38 44 39


13 1 19 23 21 30 26 23 2 29
24 15 38 34 24


Mean 1 35.8 36.8 33.2 38.2 33.5 30.7 2 37.2
38.8 35.0 39.7 41.5 33.2




Table 6. Summary of the results of the logistic regression analysis (GLIM) of the
identification scores.


Legend for Chart:


A - Subject B - Baseline C - S/N ratio, 3 dB D - S/N ratio, 6 dB E - S/N ratio, 9
dB F - S/N ratio, 12 dB G - Condition (ENH) H - Condition (ENHC10/2) I -
Interaction


A B C D E


F G H I


14 0.315 0.282 0.691 0.836 (0.078)
(0.085) (0.089) (0.091)


1.234 0.169 0.182 N.S. (0.098) (0.072) (0.072)


15 0.674 0.270 0.855 1.081 (0.083)
(0.089) (0.098) (0.103)


1.209 0.186 0.064 N.S.


(0.106) (0.080) (0.078)


16 0.816 0.442 0.627 1.157 (0.087)
(0.096) (0.099) (0.112)


1.549 -0.073 0.229 N.S. (0.125) (0.084) (0.088)


17 1.050 0.327 1.001 1.260 (0.095)
(0.101) (0.117) (0.125)


2.041 0.116 0.277 p<.01 (0.164) (0.094) (0.097)


18 1.370 0.254 0.906 1.632 (0.104)
(0.112) (0.130) (0.164)


1.385 0.258 0.359 N.S. (0.151) (0.105) (0.108)




Parameter estimates (and associated standard errors) are referenced to a
baseline, namely the control condition with 0-dB signal-to-noise ratio. A
parameter estimate for the baseline itself is given and an indication of the
significance of the interaction of signal-to-noise ratio with condition.


Table 7. Summary of the results of the logistic regression analysis (GLIM) for
the response times.


Legend for Chart:


A - Subject B - Baseline C - 3 dB D - S/N ratio (6 dB) E - S/N ratio (9 dB) F -
12 dB G - Condition (ENH) H - Condition (ENHC10/2) I - Interaction


A B C D E F G H I


14 1357 -212 -248 -350 -513 -37 -111 N.S. (29.8) (33.3)
(33.3) (33.3) (33.3) (25.8) (25.8)


15 1469 -160 -213 -293 -433 -20 -85 N.S. (25.8) (28.8)
(28.8) (28.8) (28.8) (22.3) (22.3)


16 1383 -160 -220 -310 -403 -71 -83 p < .01 (29.8) (33.3)
(33.3) (33.3) (33.3) (25.8) (25.8)


17 1535 -165 -198 -343 -432 -117 -165 N.S. (29.3) (32.7)
(32.7) (32.7) (32.7) (25.4) (25.4)


18 1378 -217 -248 -355 -512 -69 -121 p < .01 (29.3) (32.8)
(32.8) (32.8) (32.8) (25.4) (25.4)




GRAPH: Figure 1. Schematic diagram of the sequence of stages involved in the
enhancement processing of Experiment 1. The top row shows all stages of the
processing. The middle row shows the "process spectrum" stage in more detail. The
bottom row shows an example of the spectral processing for a particular frame,
for condition E3B2.


GRAPH: Figure 2. Example of the effects of the processing used in Experiment 1,
showing long-term-average spectra of a neutral vowel in noise, processed using
the control condition (NULL-top panel), condition E0 (middle panel) and condition
E3B2 (bottom panel).


GRAPHS: Figure 3. Example of the effects of the processing used in Experiment 3
showing long-term-average spectra of a neutral vowel in noise, processed using
the control condition (E0-top panel) and condition ENH (bottom panel).


GRAPH: Figure 4. Results of Experiment 3 for the second Latin Square, for the
control condition (E0), the condition involving enhancement (ENH), and the
condition involving enhancement combined with a moderate degree of compression
(ENHC10/2).


GRAPH: Figure 5. Results of Experiment 4, showing the mean (and 95% confidence
intervals for the mean) as a function of original signal-to-noise ratio and
processing condition. Panel (a) shows scores for the identification component of
the sentence verification test in terms of percent correct. Panel (b) shows the
verification component in terms of response time.


REFERENCES


1. Plomp T. Auditory handicap of hearing impairment and the limited benefit of
hearing aids. J Acoust Soc Am 1978;63:533-49.


2. Plomp R. A signal-to-noise ratio model for the speech-reception threshold of
the hearing impaired. J Speech Hear Res 1986;29:146-54.


3. Fletcher H. Auditory patterns. Rev Mod Phys 1940;12:4765.


4. Moore BCJ, Glasberg BR. Suggested formulae for calculating auditory-filter
bandwidths and excitation patterns. J Acoust Soc Am 1983;74:750-3.


5. Patterson RD, Moore BCJ. Auditory filters and excitation patterns as
representations of frequency resolution. In: Moore BCJ, editor. Frequency
selectivity in hearing. London: Academic, 1986: 123-77.


6. Moore BCJ, Glasberg BR. Formulae describing frequency selectivity as a
function of frequency and level and their use in calculating excitation patterns.
Hear Res 1987;28:209-25.


7. Glasberg BR, Moore BCJ. Derivation of auditory filter shapes from
notched-noise data. Hear Res 1990;47:10338.


8. Glasberg BR, Moore BCJ. Auditory filter shapes in subjects with unilateral and
bilateral cochlear impairments. J Acoust Soc Am 1986;79:1020-33.


9. Tyler RS. Frequency resolution in hearing-impaired listeners. In: Moore BCJ,
editor. Frequency selectivity in hearing. London: Academic Press, 1986: 309-71.


10. Dreschler WA, Plomp R. Relations between psychophysical data and speech
perception for hearing-impaired subjects. I. J Acoust Soc Am 1980;68:1608-15.


11. Dreschler WA, Plomp R. Relations between psychophysical data and speech
perception for hearing-impaired subjects. II. J Acoust Soc Am 1985;78:1261-70.


12. Glasberg BR, Moore BCJ. Psychoacoustic abilities of subjects with unilateral
and bilateral cochlear impairments and their relationship to the ability to
understand speech. Scand Audiol Suppl 1989;32:1-25.


13. Havens JF, Geisler CD. Speech recognition and frequency selectivity for
hearing impaired listeners. Proceedings of the International Conference on
Acoustics Speech and Signal Processing; 1991 May 14-17; Toronto (ON). New York:
IEEE, 1991: 3629-32.


14. Leek MR, Dorman MF, Summerfield Q. Minimum spectral contrast for vowel
identification by normal-hearing and hearing-impaired listeners. J Acoust Soc Am
1987;81:148-54.


15. Rosen S, Fourcin A. Frequency selectivity and the perception of speech. In:
Moore BCJ, editor. Frequency selectivity in hearing. London: Academic, 1986:
373-487.


16. Young ED, Sachs MB. Representation of steady-state vowels in the temporal
aspects of the discharge patterns of populations of auditory-nerve fibres. J
Acoust Soc Am 1979;66:1381-403.


17. Boers PM. Formant enhancement of speech for listeners with sensorineural
hearing loss. IPO Ann Prog Rep 1980;15:21-8.


18. Summerfield AQ, Foster J, Tyler R, Bailey PJ. Influences of formant narrowing
and auditory frequency selectivity on identification of place of articulation in
stop consonants. Speech Commun 1985;4:213-29.


19. Simpson AM, Moore BCJ, Glasberg BR. Spectral enhancement to improve the
intelligibility of speech in noise for hearing-impaired listeners. Acta
Otolaryngol Suppl 1990;469:101-7.


20. Stone MA, Moore BCJ. Spectral feature enhancement for people with
sensorineural hearing impairment: effects on speech intelligibility and quality.
J Rehabil Res Dev 1992;29(2):39-56.


21. Bunnell HT. On enhancement of spectral contrast in speech for
hearing-impaired listeners. J Acoust Soc Am 1990;88:2546-56.


22. Lim JS. Speech enhancement. New Jersey: Prentice Hall, 1983.


23. Cheng YM, O'Shaughnessy D. Speech enhancement based conceptually on auditory
evidence. IEEE Trans Sig Proc 1991;39:1943-54.


24. Clarkson PM, Bahgat SF. Envelope expansion methods for speech enhancement. J
Acoust Soc Am 1991;89:137882.


25. Allen JB. Short term spectral analysis, synthesis and modification by
discrete Fourier transform. IEEE Trans Acoust Speech Sig Proc 1977;25:235-8.


26. Patterson RD, Nimmo-Smith I, Weber DL, Milroy R. The deterioration of hearing
with age: frequency selectivity, the critical ratio, the audiogram, and speech
threshold. J Acoust Soc Am 1982;72:1788-803.


27. MacLeod A, Summerfield Q. A procedure for measuring auditory and audio-visual
speech-reception thresholds for sentences in noise: rationale, evaluation, and
recommendations for use. Brit J Audiol 1990;24:29-43.


28. Alvey N, Galwey N, Lane P. An Introduction to GENSTAT. London: Academic
Press, 1982.


29. Robinson CE, Huntington DA. The intelligibility of speech processed by
delayed long-term-averaged compression amplification. J Acoust Soc Am
1973;54:314.


30. Hecker MH, Stevens KN, Williams CE. Measurements of reaction time in
intelligibility tests. J Acoust Soc Am 1966;39:1188-9.


31. Pratt RL. On the use of reaction time as a measure of intelligibility. Brit J
Audiol 1981;15:253-5.


32. Cox RM, Alexander GC, Gilmore C. Development of the Connected Speech Test
(CST). Ear Hear 1987; 8(Suppl): 119S-26S.


33. Gatehouse S, Gordon J. Response times to speech stimuli as measures of
benefit from amplification. Brit J Audiol 1990;24:63-8.


34. Laurence RF, Moore BCJ, Glasberg BR. A comparison of behind-the-ear high
fidelity linear aids and two-channel compression hearing aids in the laboratory
and in everyday life. Brit J Audiol 1983;17:31-48.


~

By Thomas Baer, PhD; Brian C.J. Moore, PhD; and Stuart Gatehouse, PhD Department
of Experimental Psychology, University of Cambridge, Cambridge CB2 3EB, England;
MRC Institute of Hearing Research, Scottish Section, Glasgow Royal Infirmary,
Glasgow G31 2ER, Scotlan Address all correspondence and requests for reprints to:
Dr. T. Baer, Department of Experimental Psychology, University of Cambridge,
Downing Street, Cambridge CB2 3EB, England.

****** Journal of Rehabilitation Research & Development is published by VA
Prosthetics Research & Development Center and is not copyrighted.


From: Baer, et, Spectral contrast enhancement of speech in noise for listeners
with sensorineural hearing...., Vol. 30, Journal of Rehabilitation Research &
Development, 01-01-1993, pp 49.

---

Psychoacoustics - The Magic of Tone and the Art of Music - Document Three

Return to the GS WorldView Index 'Parent Directory'

---