Skip to Content

Department of Linguistics

Gap Detection Thresholds:
An Examination of Experimental Methodology

Robert Mannell
Macquarie University, 2007

Aim

This paper examines methodological issues relevant to the determination of gap detection thresholds. In order to do this we bring together the results for various experiments and examine the extent to which each of these experiments informs our development of stable and reliable methods for the examination of human auditory temporal processing.

Gap Detection thresholds are an important basic test of the temporal resolution of the human auditory system. A person's ability to hear gaps in tones and noises is related to temporal integration in the auditory system. Temporal integration is the rate at which the auditory system responds to sudden changes in intensity. Poorer gap detection thresholds often accompany hearing loss.

Readings

  1. Moore, B.C.J., 2003, The Psychology of Hearing, Academic Press, 5th edition.
    • Gap Detection: Chapter 5, esp, pp 165-176
  2. Mannell, R.H., 1994, The perceptual and auditory implications of parametric scaling in synthetic speech, Unpublished Ph.D. dissertation, Macquarie University
  3. Mannell, R.H., 2007, SPH307 Psychoacoustics lecture notes (only available to SPH307 students via WebCT on the MyMQ Portal).

Experiment 1

This experiment examines auditory temporal resolution by examining just detectable gaps in 1000 Hz tones and in broadband (white) noise.

Procedure A: Gaps in Masked Tones

When gaps are created in sinusoidal tones a spectral artifact is created ("spectral splatter"). These artifacts can be heard even when the gap is very short (eg. less than 1 ms) so it is normal to mask the tone with noise so that this additional cue to the presence of the gap is not audible. In the present experiment white noise with an average (RMS) sound pressure level about 20 dB below the level of the tone is used as the masker. At this level the spectral splatter is not audible whilst the tone is still clearly audible. Each masked tone token was 2 seconds in length.

Training sessions, similar to those used in the frequency discrimination experiment, were carried out before the final test. The first training sequence utilised tokens with clearly audible gaps of durations 100, 90, 80, 70, 60, 50, 40, 30, 20 ms. These tokens were interspersed with masked tones without gaps. A second training sequence utilised gaps with durations of 10, 9, 8, 7, and 6 ms with interspersed tones without gaps. A third and final training sequence comprised tones with very short gaps of 5, 4, 3, 2, and 1 ms, some of which were expected to be difficult to hear. Feedback, similar to that provided for the frequency discrimination experiment, was given following each training token.

The test sequence consisted of masked tones with gaps between 10 and 1 ms (inclusive) as well as tones without gaps, presented randomly. Feedback did not follow the test tokens.

Procedure B: Gaps in White Noise

In this experiment gaps were inserted into broadband (white) noise. Each noise token was 2 seconds in length. The sound intensity within each gap was effectively zero (ie. about 100 dB below the average level of the noise).

This experiment followed the same training and test procedure as for the gaps in masked tones condition.

Subjects

23 subjects participated in this experiment. The majority (about 95%) of subjects were female and their mean age was about 20. Subjects provided information about their hearing and musical background.

Results

Detailed results for the 2004 experiments are displayed in figures 4, 5 and 6.

Figure 4: This figure displays the results for the 2004 gap detection experiment in masked tones. Responses above the horizontal line represent responses that are significantly above chance (50%) for p=0.05.

You should note that relative instability (compared to the next figure) of the responses to tokens from 4 to 10 ms. This is most likely a consequence of the general tendency for noise masking to increase the degree of uncertainty in perceptual responses (as you will recall from the workshop on masked speech).

Figure 5: This figure displays the results for the 2004 gap detection experiment in white noise. Responses above the horizontal line represent responses that are significantly above chance (50%) for p=0.05.

Three features are evident in these results:-

  1. responses to a 4 ms gap are now significant (p=0.05). This suggests that tests of gap detection in broadband noise is a more sensitive measure of gap detection threshold than are masked tone.
  2. responses above 4 ms are much more stable than for the masked tones. This suggests that hearing gaps in broadband noises is more reliable than for masked tones.
  3. there is a small drop in correct response rates between 2 and 3 ms. This is possibly the result of a response bias (see below). Alternatively, it might be due to differences in the random noise characteristics at the edge of each gap which might make the 2 ms gap a bit more audible than it might be on average (tokens in the steeply sloping part of the curve are particularly sensitive to such differences).
  A response bias can occur in perception experiments as a result of many aspects of experimental design. In the present case the higher response rate for 2 ms compared to 3 ms may be biased by the fact that the 3 ms gap is preceded by a token with a 9 ms gap whilst the 2 ms gap is preceded by a token without a gap. The contrast between a preceding long and obvious gap biases subjects to not hear a following short gap.  

Chi-square statistics were applied to the above data (ie. for figures 4 and 5). The pair of 2 ms results are significantly different (p<0.001). The pair of 0 ms results are significantly different (p<0.001). The pair of 1 ms results are marginally significantly different (p<0.05).

The results for gaps ≥3 ms are significantly above chance (ie. above 50%, p<0.001). The Gap in Noise result for 2 ms is not significantly above chance (p>0.05).

Experiment 2

In experiment 2, two modifications to the experiment 1 experimental method were tried. In both experiments, the experiment 1 yes/no training procedure was repeated, but then it was followed by additional training in an AB decision task. The AB decision task presents two tokens, one of which contains the targeted change whilst the other does not change. In the simple AB decision task the subject must answer "A" if token A was perceived to change or "B" if token B was perceived to change. A response was forced, so the subject had to make a decision, even if it was only a guess. A modified version of this task, an AB0 decision task, was also attempted in a separate experiment. In this experiment subjects were told, as before, to answer "A" is they believed that token "A" changed and "B" if they believed that token "B" changed. If they didn't know which one changed they were instructed to NOT guess but to simply not respond (a zero "0" response).

23 subjects participated in both experiments.

AB selection task

Figure 6: AB gap detection in masked tone results.The curve with the filled circle symbol represents responses to token pairs in which the first token had the gap. The curve with the empty circle symbol represents responses to token pairs in which the second token had the gap. In the above diagram the horizontal line labeled "A" indicates a perfect chance response (ie. 50% A and 50% B responses). Responses equal to or greater than the line labeled"B" are significantly above chance for p=0.05. Responses equal to or greater than the line labeled"C" are significantly above chance for p=0.01.

It is immediately clear in the above figure that subjects respond very differently to token pairs in which the first token has the gap than to token pairs in which the second token has the gap. This suggests that this experiment has a much greater response uncertainty than was the case for gap detection in masked tones when the response was either "yes" or "no" (see figure 4, above). There appears to be a response bias here that favours a tone 1 response (ie. a response bias to pressing button 1 when uncertain).

Figure 7: AB gap detection in broadband noise results.The curve with the filled square symbol represents responses to token pairs in which the first token had the gap. The curve with the empty square symbol represents responses to token pairs in which the second token had the gap. In the above diagram the horizontal line labeled"A" indicates a perfect chance response (ie. 50% A and 50% B responses). Responses equal to or greater than the line labeled"B" are significantly above chance for p=0.05. Responses equal to or greater than the line labeled"C" are significantly above chance for p=0.01.

The results displayed in figure 7 show much more stable results than was the case for figure 6. The main difference between the responses for gaps in noise 1 and for gaps in noise 2 is for gap durations of 3 ms and less. This might be because there is a response bias (for pressing button 1) as a consequence of response uncertainty that might be expected for such short gaps.

AB0 Response Task

Figures 8 and 9 summarise the results for gap detection in masked tones and broadband noise using the AB0 selection task outlined above.

Figure 8: AB0 gap detection in broadband noise results.The curve with the filled circle symbol represents responses to token pairs in which the first token had the gap. The curve with the empty circle symbol represents responses to token pairs in which the second token had the gap. In the above diagram the horizontal line labeled"A" indicates a perfect chance response (ie. 50% A and 50% B responses). Responses equal to or greater than the line labeled"B" are significantly above chance for p=0.05. Responses equal to or greater than the line labeled"C" are significantly above chance for p=0.01.

The results displayed in figure 8 show an even greater response uncertainty than was the case for the equivalent AB test (figure 6). It appears that the response bias for pressing button 1 in that experiment obscured an underlying response uncertainty for gaps in tone 1.

Figure 9: AB0 gap detection in broadband noise results.The curve with the filled square symbol represents responses to token pairs in which the first token had the gap. The curve with the empty square symbol represents responses to token pairs in which the second token had the gap. In the above diagram the horizontal line labeled"A" indicates a perfect chance response (ie. 50% A and 50% B responses). Responses equal to or greater than the line labeled"B" are significantly above chance for p=0.05. Responses equal to or greater than the line labeled"C" are significantly above chance for p=0.01.

In the results displayed in figure 9, it is clear that the results are stable for gaps of 4 ms and greater regardless of which noise (1 or 2) the gap was in. The results for responses below 4 ms could be interpreted as being a consequence of a) button 1 response bias and b) the 2 and 3 ms token sequence response bias seen in figure 5 (the token order in this experiment is the same as for the figure 5 experiment).

General Conclusions

The general conclusions and observations for this series of experiments are:-

  1. Gap detection thresholds for broadband (white) noise and masked 1000 Hz tones are about 3-4 Hz but this varies a little depending upon the experimental methodology.
  2. Gap detection in masked tones appears to involve a higher degree of uncertainty than is the case for broadband noise. This may be because the broad band noise that masks spectral splatter on gapped tones reduces the certainty with which the gap can be detected. Masked tones have a higher degree of uncertainty (<100% detection) for gap widths that are reliably detected (close to 100%) in white noise. This may be because the tone has a very narrow frequency range and when the masking noise (randomly) happens to be higher at the tone frequency its increased intensity also masks the gap.
  3. Yes/no decision tasks result in less uncertainty (for gaps greater than 4 ms) than is the case for AB and AB0 selection tasks. This might be related to short term memory constraints. In the case of the yes/no decision task there is a single token and memory is not involved. In the AB and AB0 selection tasks subjects need to remember the first token and compare it to the second token, especially when the presence of the gap is uncertain.
  4. AB selection tasks appear to have a response bias in favour of selecting button 1 (A). This is true for gaps of all durations. The results suggest that when there is uncertainty subjects are more likely to select A in preference to B. This affect is much stronger for masked tones, but can also be seen for broadband noise and probably explains why there is an apparent gap detection threshold of 2 ms when the gap is in token A and a gap detection threshold of 3 ms when the gap is in token B.
  5. The AB0 selection task results display similar patterns to that of the AB selection task. The response bias for token A is particularly strong around 3 ms in the masked tone condition.
  6. 2-3 ms tokens, which are in the steepest part of the response pattern curve appear to be very susceptible to the effects of response bias, masking in both masked tones and broadband noise. These tokens might also possibly be susceptible to variations in the noise characteristics at the edges of the gap in the broadband noise which may enhance or reduce the distinctiveness of the gap.