Учебники / Hearing - From Sensory Processing to Perception Kollmeier 2007

References

Akeroyd MA, Summerfield AQ (1999) A binaural analog of gap detection. J Acoust Soc Am 105(5):2807–2820

Bernstein LR, Trahiotis C, Akeroyd MA, Hartung K (2001) Sensitivity to brief changes of interaural time and interaural intensity. J Acoust Soc Am 109(4):1604–1615

Boehnke SE, Hall SE, Marquardt T (2002) Detection of static and dynamic changes in interaural correlation. J Acoust Soc Am 112(4):1617–1626

Breebaart J, Kohlrausch A (2001) The influence of interaural stimulus uncertainty on binaural signal detection. J Acoust Soc Am 109(1):331–345

Culling JF, Summerfield Q (1998) Measurements of the binaural temporal window using a detection task. J Acoust Soc Am 103(6):3540–3553

Culling JF, Colburn HS, Spurchise M (2001) Interaural correlation sensitivity. J Acoust Soc Am 110(2):1020–1029

Dajani H, Picton T (2006) Human auditory steady-state responses to changes in interaural correlation. Her Res 219(1/2):85–100

Gabriel KJ, Colburn HS (1981) Interaural correlation discrimination: I. Bandwidth and level dependence. J Acoust Soc Am 69(5):1394–1401

Kollmeier B, Gilkey RH (1990) Binaural forward and backward masking: evidence for sluggishness in binaural detection. J Acoust Soc Am 87(4):1709–1719

Kollmeier B, Holube I (1992) Auditory filter bandwidths in binaural and monaural listening conditions. J Acoust Soc Am 92(4.1):1889–1901

Pollack I, Trittipoe W (1959) Interaural noise correlation: examination of variables. J Acoust Soc Am 31(12):1616–1618

van der Heijden M, Trahiotis C (1997) A new way to account for binaural detection as a function of interaural noise correlation. J Acoust Soc Am 101(2):1019–1022

van der Heijden M, Trahiotis C (1998) Binaural detection as a function of interaural correlation and bandwidth of masking noise: implications for estimates of spectral resolution. J Acoust Soc Am 103(3):1609–1614

van de Par S, Trahiotis C, Bernstein LR (2001) A consideration of the normalization that is typically included in correlation-based models of binaural detection. J Acoust Soc Am 109(2):830–833

Comment by Carlyon

I don’t think you can conclude that the linear relationship between LAEP amplitudes and the ‘dB(C/A) scaled IAC’ is evidence that sensitivity to changes in IAC are based on that log-ratio measure. This is because you do not have evidence that the linear LAEP amplitude is the appropriate decision statistic. For example, you could have calculated the log (or some other transform) of the LAEP amplitude, and we don’t know whether one transform is better or worse than any other (or none). Depending on the transform used, you would need to change your log-scaled ratio to some other measure in order to maintain a linear relationship. To determine what scale the LAEP should be expressed in, one would have to have measures not only the LAEP mean amplitude, but also its standard deviation. One could then use a scale where the standard deviation is constant across the range of LAEP amplitudes measured. This might be a linear scale, but it might not.

Reply

Indeed the variance over epochs in our EEG data does not depend on the IAC or the IAC transition and is roughly the same at all latencies, in particular, it does not increase at latencies corresponding to the N1-/P2-response. Since the dB(N0/Nπ) transform is the Fisher-Z-transform of the linear normalized IAC, also the variance of instantaneous dB(N0/Nπ)−IAC over time, as computed at the output of a moving temporal window, is the same for all our stimuli. Following your argument, one could easily interpret the constancy of variance of both, the EEG and the stimulus properties, in the sense that equal differences in LAEP amplitude correspond to equal discriminability by means of d-prime. However, such an interpretation would be invalid for the following reason.

In EEG recordings it is impossible to observe the activity of only those parts of the brain which are involved in binaural processing separate from other parts of the brain. Instead one always measures a far-field superposition of a comparatively tiny stimulus related brain response S (we found maximum LAEP peak amplitudes of 3–4 µV in the average over epochs) and spontaneous brain activity N which is much higher in amplitude than S and is considered as noise (with an RMS-value of about 10 µV in the filtered data before averaging).

Because the stimulus related binaural response (and in particular its possible variance) is much weaker than the spontaneous brain activity, it is not possible to separate the contributions of S and N to the EEG or its overall variance. Therefore the statistical approach described above is unfortunately not practicable for the analysis of EEG data from the auditory cortex.

Additionally, using methods similar to Shackleton et al. (2003, 2005), we computed the area below the ROC-curve corresponding to the amplitude statistics over 1000 epochs of EEG data elicited by stimuli with either the reference IAC or the deviant IAC. Even for the largest differences in IAC the ROC area (probability for correct discrimination) did not exceed 0.6, although the corresponding IAC transitions were clearly detectable in psychoacoustics (i.e., psychoacoustical ROC area close to 1).

References

Shackleton TM, Skottun BC, Arnott RH, Palmer AR (2003) Interaural time difference discrimination thresholds for single neurons in the inferior colliculus of Guinea pigs. J Neurosci 23(2):716–724 Shackleton TM, Arnott RH, Palmer AR (2005) Sensitivity to interaural correlation of single

neurons in the inferior colliculus of guinea pigs. J Assoc Res Otolaryngol 6(3):244–259

Comment by Lütkenhöner

The proposed transformation from the interaural correlation ρ to the dB-scaled ratio of diotic and antiphasic noise components, ρlog, is appealing. However, the question arises how to interpret ρlog in physiological terms. While the quantities

ρ and ρlog are more or less proportional for ρ <0.5 (roughly corresponding toρlog <5 dB), ρlog becomes infinite for ρ →1. The following consideration might help to understand this kind of singularity. The case ρ→ −1 is related to the detection of a faint diotic noise, N0, against an antiphasic noise background, Nπ. If N0 is formally considered as the signal and Nπ as the noise, ρlog may be interpreted as the signal-to-noise ratio in dB. Correspondingly, −ρlog may be interpreted as the signal-to-noise ratio related to the detection of a faint antiphasic noise against a diotic noise background. This consideration suggests that the JNDs for |ρ|≈1 might be of a quite different nature than the JNDs for |ρ|≈0 (detection threshold versus discrimination threshold).

Reply

It is correct that the basic dB(N0/Nπ)-transform as given in our article becomes infinite for values of ρ=+1 or −1. On the physiological level, however, such infinite values will never occur due to irregularities of neural processing. In our model this is simulated by adding uncorrelated noise Nu to the dichotic input signal S at a signal-to-noise ratio (SNR) of 15 dB, which is assumed to be independent of the input signal’s IAC. In a model without haircell transform, the effective “internal IAC” of the mixture (S+Nu/SNR) is then ρ(SNR2/(1+SNR2). The corresponding transformed value is 10 . log (1+ρ+1/SNR2)/(1−ρ+1/SNR2). Thus, for an SNR of 5.6 (=15 dB) the internal IAC ranges between −18 and +18 dB(N0/Nπ).

Durlach et al. (1986), Koehnke et al. (1986), Culling et al. (2001) and Boehnke et al. (2002) suggested that the BMLD and IAC-JNDs can be explained by similar mechanisms. Our model can explain both kinds of experiments, e.g., BML data by van der Heijden and Trahiotis (1998) and by Breebaart and Kohlrausch (2001). However, Culling et al. (2001) described how different cues are used depending on masker bandwidth: In broadband conditions the signal is detected as a tone in the background noise. In contrast, a signal in narrowband noise causes a percept of spatial movement of the whole stimulus. Accordingly, the issue if listeners perform a discrimination or a detection strategy could depend on the stimulus’ bandwidth rather than on the masker’s IAC.

References

Durlach NI, Gabriel KJ, Colburn HS, Trahiotis C (1986) Interaural correlation discrimination: II. Relation to binaural unmasking. J Acoust Soc Am 79(5):1548–1557

Koehnke J, Colburn HS, Durlach NI (1986) Performance in several binaural-interaction experiments. J Acoust Soc Am 79(5):1558–1562

in the LSO arrays. The CF parameter was represented along one dimension of each array according to a log-normal distribution with a mean of about 600 Hz, and ranging from 50 Hz to 1500 Hz. The best phase BP was represented along the other dimension following a Gaussian distribution with mean and standard deviation each equal to 0.3 cycles. Each value of BP was used to assign the characteristic delay: CD = BP/CF.

Responses of two model neurons to broadband noise varied in ITD are shown in Fig. 1B. For both neurons, CF = 650 Hz and BP = 0.2 cycles (CD = 308 s). The solid line represents the response of a model MSO neuron (CP = 0), and shows a peak firing rate at CD. The dashed line represents a model LSO neuron (CP = 0.5 cycles), and exhibits a null at CD.

The output of each array is the sum of its individual firing rates. The outputs are illustrated in Fig. 1D as functions of ITD for broadband noise stimulation. Each MSO is maximally active when the stimulus is in the contralateral hemisphere, while each LSO is most strongly activated by ipsilateral stimulation.

3Results

3.1Summary of Psychophysical Results to be Modeled

This section summarizes psychophysical data which illustrate a dependence of laterality on stimulus bandwidth (Trahiotis and Stern 1994), and which represent a nontrivial test of lateralization in the model. The stimulus is bandpass noise centered at 500 Hz, with ITD=1.5 ms. Figure 2A shows the pattern of activity produced by this stimulus in the BF-BD plane. The full display consists of a series of peaks and valleys, but for clarity we show only the two peaks closest to the midline (solid black lines). The straight contour is at 1.5 ms, corresponding to the ITD of the stimulus. The secondary contour is separated from the main contour by the CF period, and hence is curved.

When the stimulus is narrowband (dark gray shading), it is perceived on the left. Trahiotis and Stern (1994) explained this percept as “centrality” dominated, because it favors the secondary peak in the BF-BD plane which occurs nearer the midline. In contrast, broadband stimuli (light gray shading) are perceived on the right side. This was described as “straightness” dominated, because it favors the peak for which the ITD is consistent across CF.

Laterality also depends on the stimulus interaural phase difference (IPD). The dashed lines in Fig. 2A show the contours corresponding to a 270° phase shift. The ITD was adjusted so that the contours always passed through the center frequency at constant ITD values. Shifting the phase straightens the left contour and curves the right one. In the broadband condition, this stimulus is perceived on the left because that contour is favored by both straightness and centrality.

Fig. 2 A Contours of peak activity in CF-BD plane produced by noise with ITD=1.5 ms. Solid lines, IPD=0°. Dashed lines, IPD=270°. B Image heard to the left for narrow bandwidths and/or large IPDs. Image heard to the right for wide bandwidths

3.2Model

The dependence of lateralization on bandwidth and IPD can be explained using models that incorporate both straightnessand centrality-weighting (Stern et al. 1988). Weighting that reflects the overall tendency for physiological BD values to occur within the naturally-occurring ITD range is the most straightforward realization of centrality (Shackleton et al. 1992; Stern et al. 1988), and is an explicit property of the model described in this chapter. One method of implementing straightness-weighting is to integrate across BF along constant values of BD (Shackleton et al. 1992). This fundamentally requires a labeled-line representation because the inputs to the integration stage must be segregated according to BD. We show here that the bandwidthdependent lateralization data can also be simulated using a simple population rate code model, without explicit straightness-weighting and the resulting need for labeled lines.

Figure 3A shows the output of each of the four channels as a function of bandwidth (for IPD = 0°). We consider first a two-channel model comprising only MSO activity, and argue that it is insufficient to predict the psychophysical data. We assume that the position estimate is simply a vector sum of the MSO rates, and note that activity in one MSO corresponds to an image position in the contralateral hemifield. As bandwidth gets larger, the activity in the right MSO (RMSO) decreases (solid gray line) while the activity in the left MSO (LMSO) increases (solid black line). This correctly predicts that the lateral

Fig. 3 A Individual channel responses to noise (ITD=1.5 ms, IPD=0°) vs bandwidth. B Left and right components of model position estimates. C Model fit to psychophysical data

position moves rightward with increasing bandwidth. But the image can never actually cross to the right of the midline because RMSO is always more active than LMSO. This reflects the fact that the secondary peak produced by the stimulus is always more central than the main peak.

A four-channel model incorporating both MSO and LSO, however, can account for both the trends and magnitudes of the psychophysical data. Lateral position estimates were generated by linear combination of the channel outputs:

The parameters a = 1.55 × 10−3 and b = 1.45 × 10−3 were chosen to minimize the sum of squared error between the model position estimates and the psychophysical data. The resulting model fit (Fig. 3C) agrees well with the data (Fig. 2B), for the reasons discussed below.