Учебники / Hearing - From Sensory Processing to Perception Kollmeier 2007

Fig. 1 An example of an outcome of the loudness difference procedure. At each turning point the level differences between the variable and the variable +2 dB increase re. the reference are shown. The estimated equal loudness level is constructed from the level differences at the upper turning points of the variable and at the lower turning points of the variable +2 dB increase. The dotted lines represent the assumed loudness uncertainty region

The procedure consists out of two phases. In the first phase the limits of the auditory range are estimated by an interleaved ascending and descending stimulus sequence. In the second phase the four named intermediate categorical loudness levels are estimated. This last phase consists out of two blocks. In the first block the four named intermediate categorical loudness levels are estimated by linear interpolation between the two limits of the auditory range, which are the values at L5 (very soft) and L50 (too loud). In the second block the named intermediate categorical loudness levels are estimated by a modified least-squares fit of a linear model function. In this study three iterations of the final block were applied. In the analysis each ACALOS measurement was fitted with a model function consisting of two linear parts with independent slopes and a free cut-point. This model function is a slightly different function than the model function applied by Brand and Hohmann (2002), because their model function had a fixed cut-point at 25 CU.

2.3Subjects

In the loudness matching experiments nine normal-hearing subjects (four male, five female) participated. The age of the subjects ranged from 18 to 34 years. Two of the subjects were members of the Audiology department; the

Fig. 3 Results loudness scaling for 25 ms and 400 ms signals of different bandwidths

The figure shows that:

1.The slopes of the higher-intensity part are usually steeper than for the lowintensity part. This effect is found for all signal bandwidths and both signal durations.

2.The low-intensity slope is less steep for the 25-ms signals.

3.Furthermore, the loudness curves are ordered according to bandwidth, with a larger bandwidth leading to a higher loudness at the same level. This is what is expected from spectral loudness summation.

4.Finally, a comparison of both figures shows that corresponding levels yield a higher loudness level for the 400-ms signals than for the 25-ms signals, as would be expected from temporal integration.

In order to obtain a similar parameter of loudness summation from the loudness scaling data, we calculated for each of the stimuli the level differences relative to the level of the reference signal (60 dB), needed to obtain equal loudness as for the 400 Hz wide reference signal. Table 1 shows the calculated summation data.

At 25 ms the loudness of this signal is 11.6 CU and at 400 ms the loudness is 14.5 CU. So, there is a slight loudness differences between the reference signals

Table 1 Spectral loudness summation difference between short and long duration signals in dB SPL

at different durations. For the loudness scaling data the loudness ratings of a 800 Hz wide stimulus was used as a reference. This makes the data of the loudness matching not directly comparable to the data of the loudness scaling.

4Discussion

Although the matching procedures and the scaling procedure have been conducted with slightly different stimuli the same trends can be observed. First of all, in all three procedures spectral loudness summation is larger for short signals than for long signals. This corresponds well with the findings of Verhey and Kollmeier (2002) and Chalupper (2002). The fact that duration dependent spectral loudness summation has been found in three different measuring procedures provides extra support for the existence of this effect and excludes possible artifacts due to the measurement procedure.

The amount of spectral loudness summation difference depends on the amount of summation, which is in agreement with the results of Brand and Hohmann (2002). The duration-dependency of spectral loudness summation is small, when the loudness summation is small. As loudness summation increases, the loudness summation difference also increases. However, the maximum amount of loudness summation difference between short and long signals seems limited. In all three procedures the amount of loudness summation at a bandwidth of 3200 Hz and 6400 Hz is approximately the same. A further investigation with even broader bandwidths is needed to confirm the observation that a ceiling effect may be present. It would be interesting to determine the “critical” bandwidth at which the summation difference between long and short signals reaches the maximum value.

There are also differences between the procedures, especially with respect to the amount of summation found. This is probably a consequence of procedural differences. In the second matching procedure we assumed that interleaving of the different conditions was not necessary. Verhey (1999) found that an adaptive procedure with interleaved tracks leads to larger loudness summation. The differences we found between matching procedure 1 and matching procedure 2 correspond to the differences found between an interleaved and a non-interleaved procedure.

The long duration condition of the scaling procedure is conducted with a 400-ms signal instead of a 1000-ms signal. The influence of this difference in signal duration may be expected to be negligible, as the effect of temporal loudness integration is thought to be limited to approximately 200 ms. The scaling procedure is much less sensitive than the two matching procedures at one specific loudness. The results depend heavily on the definition of the fitting curve. Nevertheless, the results correspond reasonably well with the matching results and give also a hint towards the level dependency of the effect. At low levels there is almost no spectral loudness summation and

therefore the summation difference is also very small. Around the cut-point of the fitting curve, which seems to lie at the lower side of the most comfortable loudness region, both the spectral summation and the summation difference are largest. At higher levels they tend to decrease again.

5Summary and Conclusion

This study shows with three different measuring procedures the effect that spectral loudness summation is larger at short signal durations. Although the amount of summation differs in the different procedures the summation difference is approximately the same. Our data show a possible ceiling effect in the amount of spectral loudness summation differences between short and long signals. Further research is needed in order to investigate the effect of bandwidth on the loudness summation difference between short and long signals.

An adapted version of the model of loudness applicable to time-varying sounds (Glasberg and Moore 2002) that increases the loudness of short signals at low levels has been proven to model these effects reasonably well.

References

Brand T, Hohmann V (2002) An adaptive procedure for categorical loudness scaling. J Acoust Soc Am 112:1597–1604

Chalupper J (2002) Perzeptive Folgen von Innenschwerhörigkeit: Modellierung, Simulation und Rehabilitation. Shaker, Aachen

Glasberg BR, Moore B C J (2002) A model of loudness applicable to time-varying sounds. J Audio Eng Soc 50:331–342

Ozimek E, Zwislocki JJ (1996) Relationships of intensity discrimination to sensation and loudness levels: dependence on sound frequency. J Acoust Soc Am 100:3304–3320

Kohlrausch A, Fassel R, van der Heijden M, Kortekaas R, van de Par S, Oxenham AJ, Puschel D (1997) Detection of tones in low-noise noise: further evidence for the role of envelope fluctuations. Acust Acta Acust 83:659–669

van Beurden MFB, Dreschler WA (2005) Bandwidth dependency of loudness in series of short noise bursts. AcustActa Acust 91:1020–1024

Verhey JL (1999) Psychoacoustics of spectro-temporal effects in masking and loudness perception. PhD thesis

Verhey JL, Kollmeier B (2002) Spectral loudness summation as a function of duration. J Acoust Soc Am 111:1349–1358

Comment by Verhey

In your talk you presented a model that predicts the duration-dependent spectral loudness summation as reported in, e.g., Verhey and Kollmeier (2002, JASA 111, 1349-1358). The model was based on the assumption of a

larger gain applied to short signals than to long signals at low levels. Such a mechanism results in a higher compression at the medium to high levels. How does this assumption relate to Epstein and Florentine (2005a, b), and Anweiler and Verhey (2006) who showed that the loudness function of the short signals is essentially a vertically (downward) shifted version of the loudness function of the long signals? Such a vertical shift produces the same slope for different duration at the same intensity, i.e. no change in compression with duration.

References

Epstein M, Florentine M (2005a) Inferring basilar-membrane motion from tone-burst otoacoustic emissions and psychoacoustic measurements. J Acoust Soc Am 117:263–274

Epstein M, Florentine M (2005b) A test of the equal-loudness-ratio hypothesis using crossmodality matching functions. J Acoust Soc Am 118:907–913

Anweiler AK, Verhey JL (2006) Spectral loudness summation for short and long signals as a function of level. J Acoust Soc Am 119:2919–2928

Reply

First of all, an equal level difference at equal loudness for short and long signals over all levels is indeed in disagreement with my hypothesis. A small decrease in level difference at equal loudness is expected at low levels. In fact the data of Epstein and Florentine (2005b) is somewhat ambiguous and a small decrease in level difference at equal loudness can be seen. If the lowest point is neglected a decrease in loudness ratio at low levels is found. The ambiguity can also be seen in the individual data in which some subjects seem to show a clear level difference decrease at low levels (L5,L6,L9) In that case no contradiction between the adapted model and their data is present. The magnitude estimation data of Epstein and Florentine (2006a) show a clear increase in level difference at equal loudness at low levels, which is clearly in contrast with my hypothesis. But here the group data appear to be heavily influenced by subject L3 and the individual data again show also subjects with a decrease in level difference at low levels (L5, L6). The scaling data from Anweiler and Verhey (2006) and the present paper is not accurate enough at low levels to do well funded statements. Therefore I think it is not possible to say if these data agree, or disagree with my hypothesis.

Unfortunately Epstein and Florentine (2005a, b) have not presented data for broadband signals. For such signals the model predicts a larger decrease in level difference at low levels than for narrowband signals, which should be easier to measure.

frequency modulations, sound timbre and spatial location as the most prominent candidates for grouping cues in auditory streaming (Cooke and Ellis 2001). It is, however, becoming more evident that any sufficiently salient perceptual difference along any auditory dimension (at the periphery or central auditory stages) may lead to stream segregation.

On the biophysical level, our knowledge of neural properties particularly in the auditory cortex indicates that cortical responses (Spectro-Temporal Receptive Fields, STRFs) exhibit elaborate selectivity to spectral shapes, symmetry and dynamics of sound (Kowalski et al. 1996; Miller et al. 2002). This intricate mapping of acoustic waveforms into a multidimensional space suggests a role of the cortical circuitry in representing sounds in terms of auditory objects (Nelken 2004). Moreover, this organizational role is supported by the correspondence between time scales of cortical processing and the temporal dynamics of stream formation and auditory grouping.

In this study, we formalize these principles in a computational scheme that emphasizes two critical stages of stream segregation: (1) mapping sounds into a multi-dimensional feature space; (2) organizing sound features into temporally coherent streams. The first stage captures the mapping of acoustic patterns onto multiple auditory dimensions (tonotopic frequency, spectral timbre and bandwidth, harmonicity and common onsets). In this mapping, acoustic elements that evoke sufficiently non-overlapping activity patterns in the multi-dimensional representation space are deemed perceptually distinguishable and hence may potentially form distinct streams. We assume that these features are rapidly extracted and hence this mapping simulates “instantaneous” organization of sound elements (over short time windows; e.g. <200 ms), thus evoking the notion of simultaneous auditory grouping processes (Bregman 1990).

The second stage simulates the sequential nature of stream segregation. It highlights the principle that sound elements belonging to the same stream tend to evolve together in time. Conversely, temporally uncorrelated features are an indication of multiple streams or a disorganized acoustic scene. Identifying temporal coherence among multiple sequences of features requires integration of information over relatively long time periods (e.g. >300 ms), consistent with known dynamics of streaming-buildup. Therefore, the current model postulates that grouping features according to their levels of temporal coherence is a viable organizing principle underlying cortical mechanisms in sound segregation.

2.2Stage 1: Multi-dimensional Cortical Representation

Current understanding of auditory cortical processing inspires our model for the multi-dimensional representation of sound. The model takes in as input an auditory spectrogram, and effectively performs a wavelet decomposition using a bank of linear spectro-temporal receptive fields (STRFs). The