Experiment 1 was designed to explore how the threshold for detecting a narrow band noise alters as the properties of a number of flanking noise bands are altered. The conditions were chosen so that comparisons of subsets of the results would indicate the relative importance of flanking bands close to or far away in frequency from the signal band. This was done to get a preliminary indication of the relative roles of within- and across-channel cues.
A single signal band of noise and up to six masking bands were used in the present experiment. The signal band was always centred at 1500 Hz. The masking bands were also fixed in centre frequency. The bands were spaced so they were roughly the same distance apart on an ERB scale (Glasberg and Moore, 1990). The separation of each of the bands was chosen so that they were roughly 2 ERBs apart. This was done so that to a first approximation each band would be in an independent frequency channel.
The envelope of each masking band could be correlated or uncorrelated with that of the signal band. Alternatively, a given flanking band could be replaced by a sinusoid with the same average power or be absent. If masking bands were uncorrelated with the signal, they could be correlated with each other (co-uncorrelated) though this was not necessarily the case. Of course, with such a large array of manipulations and with multiple bands, there is a prohibitive number of possible conditions. 14 conditions were chosen which are graphically demonstrated in figure 2.1.
Figure 2.1. The 14 different conditions used in Experiment 1. The centre white block represents the signal at 1500 Hz. The other boxes represent the bands of noise that make up the masker (547, 765, 1071, 2100, 2940 & 4116 Hz). Similarly shaded blocks are comodulated. A gap indicates that no band is present at that frequency. A letter 'T' indicates that a single sinusoid is present instead of a band of noise.
The conditions were chosen for the following reasons:
Condition 1.
All bands are correlated, thus providing a baseline condition.
Condition 2.
All the flanking bands are correlated with each other. The signal band is uncorrelated. This also provides a baseline condition and when taken in conjunction with condition 1 should provide a demonstration of CDD with a large number of flanking bands.
Condition 3.
All the bands are correlated except for the 2 bands that are placed nearest to the signal. This tests whether CDD depends only on the most proximal bands or whether the correlation of more distant bands also plays a role (by comparison with condition 2).
Condition 4.
This is almost the reverse of condition 3. All the flanking bands are uncorrelated with the signal except for the 2 bands that are nearest to the signal. Comparison of condition 4 to the baseline conditions 1 and 2 will indicate the relative contributions of correlation of the proximal and distal bands.
Condition 5.
Only the 2 innermost bands are present as well as the signal band. All bands are correlated. This provides a baseline condition for the innermost bands being correlated.
Condition 6.
Only the 2 innermost bands are present as well as the signal. The 2 flanking bands are correlated with each other, but the signal is uncorrelated. This provides a baseline condition for the innermost bands being uncorrelated with the signal. In conjunction with condition 5, this should provide a demonstration of CDD with a small number of flanking bands.
Conditions 7 and 8.
These condition are intended for comparison with condition 3. All the bands are correlated, except for one pair in each case. This should show how moving a pair of uncorrelated bands away from the signal in a background of correlated bands alters the threshold.
Condition 9.
This condition is like condition 5 except that the 2 correlated bands are farther away from the signal.
Conditions 10 and 11.
In condition 10, all the bands are completely independently modulated. In condition 11, the bands are split into pairs based on distance from the signal. The 2 halves of each pair are correlated with each other, but each pair is independent. By comparison with condition 2, the importance of the correlation among the flanking bands can be assessed.
Condition 12.
All bands are co-uncorrelated with the signal except for the outermost bands which are correlated with the signal. This is to test whether the outermost bands have any effect on the threshold. If so, this would almost certainly reflect an across-channel process.
Conditions 13 and 14.
These condition are intended for comparison with conditions 5 and 6. Instead of the 2 outer pairs of bands being missing, they are replaced by unmodulated sinusoids. This is to test a) whether unmodulated bands have any effect and b) whether the modulation pattern of energy farther away from the signal frequency is important.
Pilot studies showed that signal threshold in condition 9 was not significantly above absolute threshold. Therefore it was omitted from the experiment proper. It should be noted that the condition number bears no relation to the condition that it describes. The conditions have not been arranged so as to form a trend.
Both the signal and maskers were composed of 20 Hz wide bands of noise generated digitally using either a Masscomp 5400 or a Silicon Graphics Indy computer. Two 'tracks' were generated. The signal was put on one track and the maskers on another. Each 20-Hz wide band of noise was produced by adding together 201 sinusoids spaced 0.1 Hz apart, each with a random phase and amplitude (being drawn from a rectangular and Rayleigh distribution, respectively). Each band of noise was therefore a harmonic series with a 0.1 Hz fundamental. One full period (10 seconds) of each track was calculated, after which the track could be looped back to the beginning, to create a continuous noise.
To make two bands correlated, the phase and amplitude of each component in one band was the same as for the analogous component in the other band. For example, the 1st components in each band had the same phase and amplitude, even though they had different frequencies. The 134th components (say) had the same phase and amplitude as each other, but this was unrelated to the phase and amplitude for any other component. This is illustrated pictorially in figure 2.2.
Figure 2.2. This shows how masking bands can be constructed to be correlated or uncorrelated with the signal band. Only 10 components of the 201 in each band are shown. The height of each bar corresponds to intensity or starting phase.
The signal and masker tracks were adjusted so that the signal band was at a level 10 dB higher than that of each of the masker bands. The signal and masker were played out simultaneously through twin 16-bit digital-to-analogue converters (Masscomp DA04H) at a sampling rate of 16000 Hz. They were low-pass filtered at 7 kHz using Kemo VBF8 filters (96-dB/oct slope) and then recorded onto separate channels of a digital audio tape (Sony 55ES). The tape was then played continuously for each condition.
The signal was a 20Hz wide band of noise centred at 1500 Hz gated with 50ms raised-cosine ramps and with a steady-state duration of 300 ms. The start of each trial was initiated by the subject's response on the previous trial. As the subjects did not have a fixed response time, the timing of the 300 ms presentation was, in effect, randomly chosen from the continuously looping signal and masker tracks. Therefore, there was no possibility of subjects learning the modulation patterns.
The signal was presented at random in one of three observation intervals which were marked by lights on the response box. The intervals were separated by 200 ms. Signal timing was controlled by a Texas Instruments 990/4 computer. Two analogue multipliers (AD534L) connected in series were used as a gate. The gating voltage was derived from a 12-bit digital-to-analogue converter. The signal level was controlled by a Charybdis model D programmable attenuator.
The masker was gated simultaneously with the three intervals in which the signal might appear and consisted of up to six bands with centre frequencies of 547, 765, 1071, 2100, 2940 and 4116 Hz. The centre frequencies were chosen to be approximately two ERBs apart from each other. In two of the conditions, all bands except for the innermost two (1071 & 2000) were replaced by single sinusoids. Each sinusoid had the same overall level as one band of noise (i.e. its level was 13 dB higher than the spectrum level of the noise). The masker and attenuated signal bands were added together and passed through another programmable attenuator, in order to control the overall level. The stimuli were then passed through a manual attenuator before being delivered to the left ear piece of a Sennheiser HD 414 headset. The final spectrum level of the maskers was either 50 dB SPL or 65 dB SPL.
Masked thresholds were estimated using a three-alternative forced-choice method with an adaptive three-down one-up procedure estimating the 79.4% point on the psychometric function. The initial step size was 5 dB. After three reversals, the step size was decreased to 2 dB. A run was terminated after 12 reversals. The threshold was defined as the mean of the levels at the last eight reversals. Three thresholds estimates were obtained for each condition, and a fourth was obtained if the range of the first three exceeded 3 dB. The final threshold value was taken as the mean of these three (or four) estimates. Subjects were tested individually in a double-walled sound-attenuating booth. Feedback was provided on the response box.
Three subjects participated, all with absolute thresholds less than 10 dB HL at all audiometric frequencies. One subject was the author, who had extensive practice during pilot tests. The other two subjects were paid volunteers, who were given several hours of practice. Subject SB and GEM carried out the experiment twice, once with masker spectrum levels of 50 dB SPL and once with spectrum levels of 65 dB SPL. Subject GPM only carried out the experiment with masker spectrum levels set at 65 dB SPL.
The results for the 50 dB SPL spectrum level condition are presented in figure 2.3 (Experiment 1a). Table 2.i shows the average results for the two subjects. Table 2.ii shows the results of a pairwise comparison between all the conditions averaged across subjects. Most analysis was performed on the results of the 50 dB SPL spectrum level condition; however, the results for the 65 dB SPL condition (Experiment 1b) are very similar in form. Experiment 1b was actually carried out after later experiments which used the higher level maskers. The results of experiment 1b are presented in this section for completeness.
Figure 2.3. The results of experiment 1a for 50 dB SPL spectrum level masker bands. A reminder of the conditions is also presented. The error bars show the standard deviation of the results.
Due to the large number of conditions in this experiment, the best way to interpret the results is to compare a series of small groups of conditions. Each group is chosen so that one particular manipulation is investigated. Some of the most important groups are discussed below.
Comparison A.
5 vs 6 and 1 vs 2.
This is a demonstration of CDD with 2 and 6 masking bands. All the masking bands are correlated with each other and the correlation of the signal is altered. With 2 masking bands the CDD is 4.9 dB. When the number of masking bands is increased, the CDD rises to 10.4 dB. This fits in with an across-channel grouping point of view as it is assumed that if there are more bands that are correlated (i.e. the flankers which are all correlated with each other in all conditions), any bands that are uncorrelated (i.e. the signal) will stand out more. This is also consistent with an across-channel dip listening theory as increasing the number of masker bands produces more channels providing information on the optimum times to listen for the signal in the co-uncorrelated condition. However, the inverse comparison (B) gives:
Comparison B.
1 vs 5 and 2 vs 6.
These conditions demonstrate the effect of adding more flanking bands of the same type. If the bands are correlated, the threshold rises from 24.3 dB to 29.3 dB. If the flanking bands are uncorrelated with the signal, the threshold decreases slightly (non-significantly) from 19.4 dB to 18.9 dB. These comparisons show that the difference in CDDs observed in comparison A is mainly due to an increase in threshold produced by adding more correlated bands rather than to a decrease produced by adding more uncorrelated bands. This is demonstrated by there being no significant difference between conditions 2 and 6. An across-channel mechanism would predict that increasing the number of co-uncorrelated bands would decrease the threshold (i.e. detection would be facilitated). This is because more channels would be involved in the comparison process. The across-channel grouping mechanisms described above could predict more masking with an increasing number of correlated bands which is indeed observed. The effect of combination bands is discussed in the next comparison. Further evidence for across- and within-channel explanations is given in the other comparisons.
Comparison C.
1, 4, 7, 8 and 13 vs 5.
Conditions 1, 4, 7, 8 and 13 all have 2 correlated inner masking bands. The outer 4 bands are either replaced by non-modulated sinusoids (condition 13) or are every combination of correlated and uncorrelated bands (subject to the restriction that they are symmetrically modulated, e.g. the outermost bands are always correlated with each other). These conditions all give very similar thresholds (except for 7 which has a slightly higher threshold), but they all give higher thresholds than condition 5 in which the outer 4 bands are missing. The mean increase is 5.2 dB. Therefore, the modulation characteristics of bands centred at frequencies away from the signal are not important, but the presence or absence of the bands is important. The dominance of the innermost bands could be caused by either within- or across-channel mechanisms.
The overall level of the maskers increases by 4.8 dB when 4 extra bands are added, however they are modulated. This gives rise, on average, to an extra 5.2 dB of masking. As the more distant bands will be attenuated markedly by the auditory filters, this rise is surprisingly high. A likely explanation is connected with combination bands produced by non-linearities in the ear (Greenwood, 1971). When there is only one band lower in frequency than the signal, a cubic combination band (2fm-fs i.e. at 2x1071-1500=642Hz) would probably be detectable. When there are 3 bands below the signal frequency such a combination band would be much less detectable due to the proximity of the 547-Hz band. It is not clear what advantage would be gained by listening for a combination band which had maskers roughly 1 ERB on either side compared to listening for the signal band itself which was placed 2 ERBs away from the masker bands. The combination band would also be at least 30 dB lower in level than the signal. There is also the possibility that combination bands produced by interactions between the upper masking bands would mask the signal (the 2100, 2940 and 4116 Hz bands would produce combination bands at 1260 and 1764 Hz). Such bands could be as high in level as 30 dB down from the source bands. i.e. 50 dB spectrum level maskers would produce 20 dB spectrum level combination bands, corresponding to 33 dB SPL overall.
Comparison D.
2, 3, 12 and 14 vs 6.
Performing the analogous set of comparisons to comparison C, but with the innermost bands uncorrelated, produces an entirely different pattern of results. Conditions 2, 12, 14 and 6 now give very similar thresholds; condition 3 gives a threshold some 5.3 dB higher. It appears that in condition 3 the correlated bands adjacent to the inner bands are responsible for some of the masking. Comparisons C and D suggest that the closest bands are responsible for all the masking if they are correlated. This could be seen to be at variance with the large difference between condition 5 and the other conditions in comparison C. However, the difference seen in comparison C may be largely due to combination bands. Certainly, one must be cautious of comparisons with condition 5 as the thresholds for that condition probably reflect detection of combination bands. Such bands were probably not detectable in condition 6 because thresholds were so low that the bands would have been below absolute threshold. Also, combination bands will generally be less intense when the source bands are uncorrelated. If the closest bands to the signal are not correlated, then some extra masking may be produced by the next set of bands but only when they are correlated with the signal band. Note that this does not extend to the outermost bands. This is demonstrated by the threshold for condition 12 being no higher than that for condition 2. There is no clear reason why the outermost bands should have no effect from an across-channel grouping view, unless grouping is less likely if the components are widely separated in frequency. However, it is easily explained by a within-channel approach as components produce much less output from a given auditory filter (centred on the signal band) as they move away from its centre frequency. One way to test this would be to use higher level maskers. As filter bandwidth increases with level (Lutfi and Patterson, 1984; Moore and Glasberg, 1987), one would expect more extreme frequency regions to have a greater effect with a greater input level. This is especially true for low frequency regions (the Upward Spread of Masking). Higher level maskers were used in experiment 1b.
The difference between conditions 3 and 12 is consistent with both within- and across-channel mechanisms. There are more bands comodulated with the signal in condition 3. Therefore, an across-channel grouping mechanism would predict that signal would be more tightly grouped with the masker. Having more modulated bands decreases the number of channels which can be usefully monitored for dip-listening or correlation cues. A within-channel mechanism would predict that the summed masker envelope at the output of a filter centred on the signal freqency would be flatter due to the summation of bands that are not correlated with each other. Therefore, the within-channel envelope flattening cue will be diminished. However, given that the middle bands (centre frequencies of 765 and 2940 Hz) are nearly one octave apart from the signal and that there are bands between them and the signal, it is unlikely that they would have much of a within-channel effect. Notwithstanding, the difference between conditions 3 and 12 as the spectrum level is increased (see experiment 1b) is consistent with a within-channel mechanism.
The results of experiment 1b (which was a repeat of experiment 1a, but with 65 dB SPL maskers are shown in figure 2.4 and Table 2.iii. The results generally show a very similar pattern to those in experiment 1a. The thresholds are approximately 25 dB higher than those in experiment 1a. The masker levels were only 15 dB higher, which is consistent with the upward spread of masking.
Figure 2.4. The results of experiment 1b for 65 dB SPL spectrum level masker bands. A reminder of the conditions is also presented. The error bars show the standard deviation of the thresholds
TABLE 2.iii. This table shows the detection thresholds for a 20Hz wide band of noise centred at 1500 Hz with a variety of different masker configurations. The masking bands were at 65 dB SPL spectrum level. The results are the average for three subjects.
Compared to experiment 1a, the results show a large overall difference in threshold, but the same patterns are seen. The comparisons detailed above generally hold true at the higher level. If an across-channel mechanism is used, the difference between conditions should not be affected by a change in level (as long as the change is not too great), though the thresholds will increase, of course. By comparing the mean results of experiments 1a and 1b, the within-channel effects of the upward spread of masking can be investigated. Figure 2.5 shows the average results from both experiments. The average for experiment 1b has been calculated from the results of subjects SB and GEM as only these two subjects took part in experiment 1a. The average difference in thresholds between the two noise levels is 25.2 dB. The results of the experiment 1a (50 dB spectrum level) have all been raised by 25.2 dB, to permit an easier comparison of the conditions.
Figure 2.5. A comparison of the results of experiment 1a and experiment 1b. As the thresholds for the lower level condition are lower, they have been raised by the average difference (25.2 dB). The results are the average for subjects SB and GEM. A reminder of the conditions is also presented. The error bars show the standard deviation of the thresholds.
The most important difference in the results between the two masker levels is the change in condition 3:
Condition 3 has a much lower threshold than condition 1 at low masker levels. At high masker levels, the thresholds are almost the same. This suggests that the correlation of the outer bands is having a greater effect at the higher level. The upward spread of masking would mean that the correlated bands (which increase thresholds) produce a much greater output from the filter centred on the signal. If the proximal bands are correlated (as in condition 1), then they will be responsible for a large amount of masking. If the proximal bands are uncorrelated (as in condition 3), the correlation of the bands further away from the signal frequency will have a greater effect. This does not extend to the most distal bands, however. The results for condition 12 are almost the same as those for condition 2 at both levels. By comparing conditions 2 and 3 at both levels, the greater effect of the outer correlated bands can be seen.
The masked threshold of the signal can be viewed as being made up of two components; a modulation-dependent component (MDC) and a modulation-independent component (MIC). The MIC is probably due to the long-term power of the maskers at the output of the filter centred on the signal frequency. As the masker level rises, the filter broadens and so the output of the filter changes by a greater amount than the rise in masker level. Therefore, the MIC will rise by a greater amount than the rise in masker level. The average increase in threshold as the level increases is due to an increment in the MIC. In experiment 1, the increase in the MIC is 25.2 dB for a 15 dB rise in masker level. Any change in the results for one condition over and above the average MIC should be assigned to the MDC for that condition. At first, it is not clear why should there be a change in the MDC as the level rises. If the CDD is due to an across-channel process, there should be no difference at all, as the relative contributions of the channels are not altered. However, a within-channel mechanism would predict a change in some conditions, especially those in which inner bands are uncorrelated and outer ones are correlated (conditions 3 and 12 in experiment 1). The increase in filter width at higher levels would mean that there would be more masking from more distant correlated bands.
As the comparisons detailed earlier suggest that CDDs are due to correlated masking bands increasing the threshold, rather than uncorrelated bands decreasing the threshold, the MDC could therefore be defined to be zero when all the bands are co-uncorrelated. The MDC could also be defined to be the increase in threshold due to correlated flanking bands. This allows the relative contribution of the MIC and MDC to be calculated. There might appear to be a problem with this description however, as thresholds differ between conditions 2, 10, 11 and 14.
In these conditions, all the flanking bands are uncorrelated with the signal band. In condition 10 all the flanking bands have independent envelopes. In condition 11, there are three pairs of correlated bands. In condition 14, the outer two pairs of bands have been replaced by sinusoids. Condition 2 gives the lowest threshold. The output of a filter centred on the signal (when the signal was not present) would have an envelope with almost the same statistics as the individual masker bands. The rate of fluctuations would be the same. The peaks of each masker band would reinforce each other. In conditions 10, 11 and 14 this is not the case. The peaks in a given band would tend to be out of phase with the peaks in other bands. Peaks would tend not to reinforce each other. Indeed, they would tend to fill troughs. Hence, the waveform would be flatter and the rate of fluctuations would increase. An across-channel mechanism based on dip-listening would not be as successful as there would be fewer dips. A within-channel mechanism may use the cue of envelope flattening to detect the signal in a co-uncorrelated background of maskers. As the envelope is already smoothed when the maskers are not correlated with each other, this within-channel cue will be impoverished. This would lead to the rise in threshold observed. As this rise is modulation-dependent, it should be part of the MDC. The MDC is therefore defined to be the increase in threshold over that for the co-uncorrelated condition with the same number of masking bands.
The threshold when the masker bands are not correlated with each other (or with the signal band) is greater than that produced by co-uncorrelated bands (condition 2). It is therefore necessary to resolve the MDC into two components; an interference dependent component (IDC) caused by the disruption of the within-channel envelope flattening cue (as detailed above; this is related to the lack of correlation of masker bands with each other) and a number of correlation dependent components (CDCs) corresponding to each pair of masking bands. The CDCs are due to correlated flanking bands being more effective maskers. The analysis of the results should therefore split each threshold into at least three components; the MIC, the IDC and between 1 and 3 CDCs. The analysis for experiment 1b is described below. The terminology used is as follows: MICi is the modulation independent component of the threshold when there are i masking bands. CDCP is the correlation dependent component for a pair of bands in position P (where P is equal to I for inner, M for middle, O for outer pairs or ALL for all masking bands). IDCj is the interference dependent component for condition j.
º MIC2+IDC6 = 41.6 dB
IDC6 = 0 as the masker bands are correlated with each other
Þ MIC2 = 41.6 dB
º MIC2+IDC5+CDCI = 48.1 dB
IDC5 = 0 as the masker bands are correlated with each other
\ 41.6+0.0+CDCI = 48.1 dB
Þ CDCI = 6.5 dB
º MIC6+IDC2 = 43.6 dB
IDC2 = 0 as the masker bands are correlated with each other
Þ MIC6 = 43.6 dB
º MIC6+IDC12+CDCO = 43.5 dB
\ 43.6+IDC12+CDCO = 43.5 dB
IDCs and CDCs are assumed to be positive.
Þ IDC12 » 0.0 dB
CDCO » 0.0 dB
º MIC6+IDC1+CDCI+CDCM+CDCO = 52.3 dB
IDC1 = 0 as the masker bands are correlated with each other
\ 43.6+0.0+6.5+CDCM+0.0 = 52.3 dB
Þ CDCM = 2.2 dB
It can be seen that the CDCs decrease as the correlated bands get further from the signal frequency (6.5 dB®2.2 dB®0.0 dB). This is entirely consistent with a within-channel explanation as there will be a greater attenuation of the masking bands that are at increased frequency separations from the filter centred on the signal frequency. This analysis does not take combination bands into account. It has been suggested that combination bands would be masked in conditions with six masking bands and hence only conditions 5 and 6 might have been affected. Furthermore, as the threshold in condition 5 was so low and combination bands would be less intense with uncorrelated bands, it is likely that combination bands probably only affected the thresholds in condition 6. The combination bands would have the effect of reducing the threshold. If we define the Combination Tone Component (CTC) to be an unknown positive value in condition 5 and zero in all other conditions, then the affected parts of the previous analysis can be adjusted:
º MIC2+IDC5+CDCI-CTC5 = 48.1 dB
IDC5 = 0 as the masker bands are correlated with each other
\ 41.6+0.0+CDCI-CTC5 = 48.1 dB
Þ CDCI = 6.5dB + CTC5
º MIC6+IDC1+CDCI+CDCM+CDCO = 52.3 dB
IDC1 = 0 as the masker bands are correlated with each other
\ 43.6+0.0+6.5+CTC5+CDCM+0.0 = 52.3 dB
Þ CDCM = 2.2 dB - CTC5
Therefore, increasing CTC5 has the effect of increasing the estimated dominance of the inner bands over the middle bands even more, which is still entirely consistent with the within-channel explanation as described above. The analysis also predicts that the maximum CTC can only be 2.2 dB (as the CDC values are assumed to be positive).
Now that all the CDCs are known, the IDCs can be easily calculated:
| Condition | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 10 | 11 | 12 | 13 | 14 |
| IDC | 0.0 | 0.0 | 7.1 | 5.5 | 0.0 | 0.0 | 7.0 | 2.3 | 3.4 | 5.5 | 0.0 | 3.9 | 5.6 |
The IDCs are very similar in magnitude to the CDCs and are also affected by the greater attenuation of the more distal bands. In the conditions where the masking bands are not all correlated with each other (and hence are not expected to have a zero IDC), the smallest IDCs are found in conditions 8 and 12. In these conditions the outermost bands are uncorrelated with the rest of the masking bands. As the outer bands are attenuated greatly by the filter centred on the signal frequency, they do not decrease the availability of the within-channel envelope smoothing cue. Hence, the IDCs are very small. Larger IDCs are seen when the inner and middle pairs of bands are not correlated.
The technique of splitting the masked threshold into a number of additive components assumes that the components are independent. For instance, if a pair of correlated bands is moved farther away from the signal frequency, the contribution of CDC is reduced. This can also alter the IDC. However, the IDC is defined in terms of the extent to which correlated bands are unable to account for the whole threshold. This means that the splitting method described here forces the components to be additive at the expense of some circularity in the definitions.
Is it fair to assume that the components of the masker could all increase the masked threshold in an additive manner? For example, if the masker consisted of a square wave 100% modulated at 1 Hz, then threshold for detection of a continuous signal would be roughly at absolute threshold. If more maskers were added at different frequencies, but similarly modulated and in phase, then the masked threshold would not increase. However, this is very different from the stimuli used in experiment 1 in which the maskers have a much faster modulation rate, the modulation is random and the modulation depth is random. More importantly, the signals are modulated. If in the square wave example, the signal was modulated at the same rate as the maskers, but was either in-phase or anti-phase the results would indeed be roughly additive. The anti-phase signal would have a masked threshold at absolute threshold due to detection in the silent gaps. The in-phase signal would be detected in the same manner as a continuous signal with continuous maskers. There would be no across-channel processes used in detection.
Experiment 1 provides an adequate demonstration of CDD. CDD appears to be caused by correlated bands giving rise to elevated thresholds rather than uncorrelated bands giving decreased thresholds. Correlated bands close in frequency to the signal have a much larger effect than bands further away. Indeed, if the innermost bands are correlated, then the modulation patterns of bands at more remote centre frequencies are unimportant. If the innermost bands are uncorrelated, then the other bands become more important. Correlated bands further away in frequency will increase threshold.
Unfortunately, like many other CDD experiments, the experiment detailed here suffers from having too many conditions. This makes evaluation of the results difficult. By comparing a small carefully chosen number of conditions it is still entirely possible to draw valid conclusions, however.
As the bands were all fixed in frequency, it is not easy to investigate why correlated bands close in frequency to the signal have such a large effect. It is easy to explain from a within-channel point of view. However one could also postulate that there is a 'sweet spot' (or more sensitive region) for grouping which exists at a separation of between 30% and 40% of the signal frequency, say. Wright (1994) reviewed experiments which showed that maskers with frequency components presented about 1.2 times the signal frequency (i.e. higher in frequency) provided signal enhancment in across-channel tasks. The 30-40% spacing in the current experiment would be roughly 2 ERBs from the signal and so would classically be viewed as being in an independent frequency channel. Auditory filters are not rectangular and at higher levels especially, there could be a significant output from the filter centred on the signal frequency from masker bands at remote frequencies. Testing such a hypothesis would require the bands to be altered in frequency. This is addressed in later experiments.