Bruce G. Berg

Analysis of weights in multiple observation tasks

Multiple observation tasks are often used to investigate observers’ abilities to combine information from different sources or observations. This paper discusses a method for assessing the relative weight given to each observation. First, a theorem that derives the relative weights is stated. A technique for estimating relative weights is then illustrated by analyzing data from a multiple‐tone task. In this task, observers listen to a sample of tones drawn from one of two distributions and decide which distribution was sampled. Results show that observers integrate information nonoptimally, attaching greater weight to the first and/or last observations than to observations occurring in the middle of the sequence. The generality of the technique is discussed.

Observer efficiency and weights in a multiple observation task

A sequence of seven tones, sampled from one of two distributions differing in mean frequency, is presented to observers who try to report which distribution was sampled. Estimates are obtained of the weight or importance given to each tone as a function of its temporal position. In five experiments, the reliability of the information is varied by changing the variance of the distributions; tones with high reliability are sampled from distributions with relatively small variance, whereas tones with low reliability are sampled from distributions with relatively large variance. Results show that the observations are weighted more efficiently when the tones have equal rather than unequal reliability, and when the most reliable tones have the greater intensity. Additional results show that the most intense tones often receive the greatest weight, even when those tones have the least reliability.

Spectral weights in profile listening

The COSS analysis suggested by Berg [B. G. Berg, J. Acoust. Soc. Am. 86, 1743-1746 (1989)] is applied to a profile listening task. The listeners task is to detect an increment in the level of the middle component of an n-component spectrum. The overall level of the components is randomly selected from a 20-dB range on each presentation; thus the detection task is essentially one of detecting a change in spectral shape. To implement the COSS analysis, a small perturbation in level is added to each component of the complex. COSS functions are generated from these perturbations, and the spectral weight that the listener assigns to each component is estimated. Data are reported for n = 3, 5, and 11 components and for perturbations with standard deviations of 0.5, 1, and 2 dB. The estimated weights are similar to those derived for an optimum detector; namely, the level at the signal component is compared with the average level of the nonsignal components. This result supports the view that profile analysis involves an across-channel comparison process. The pattern of weights also provides insight into differences among listeners. In a separate experiment, the spectral weights of a very poor profile listener are estimated, and the pattern of the weights suggests reasons for the inferior detection performance.

On the relation between comodulation masking release and temporal modulation transfer functions.

Explanations for the phenomenon known as comodulation masking release (CMR) generally assume that temporal envelope information from different peripheral filters is compared, thus entailing multiple envelope representations. Here it is shown that a leaky-integrator model, yielding a single envelop representation extracted from a broad frequency range, provides an alternative account of CMR. One prediction obtained from model simulations is that adding a single tone to the stimulus will disrupt the CMR effect, leading to an increase in thresholds. Supporting evidence from several experiments shows that the magnitude of CMR is reduced following the addition of a single tone, even when separated from comodulated masking and flanking bands by more than an octave. Whereas these findings are consistent with the leaky-integrator model, they cannot be easily explained by models which assume multiple envelope representations.

A molecular description of profile analysis: Decision weights and internal noise

Systematic inefficiencies and internal noise in a spectral profile discrimination task were investigated. Listeners detected a 1000-Hz sinusoid added in-phase to the central component of a complex consisting of 11 equal-intensity sinusoids. Parameters for a channel model that employs decision weights and internal noise were estimated with molecular psychophysical techniques. Maximum likelihood predictions of the model were generally within a few decibels of observed thresholds. The degree to which an assumption of ideal weights leads to overestimation of internal noise was also assessed.

Spectral and temporal weights in spectral-shape discrimination.

The COSS analysis [B. G. Berg, J. Acoust. Soc. Am. 86, 1743-1746 (1989)] was used to estimate spectral and temporal weights of a three-component, amplitude-modulated stimulus in a spectral-shape discrimination task. In all experiments, the task of the observer was to detect an increment in the level of the center component. A spectral-temporal weight quantifies the relative influence of a spectral component on the decisions of an observer during a specified segment of the total stimulus duration. In the first two experiments, the signal was added to all three temporal segments of the center component. The ideal weights for each component should be the same across temporal segments. Spectral-temporal weights were obtained for four conditions with different stimulus durations. In general, the estimated weights for each component were not equal at different temporal segments. In the third experiment, the signal was added to only one of three segments of the center component. Ideally, weight patterns should have changed when the temporal position of the signal segment was altered. Two stimulus durations, 300 and 15 ms, were used. For the 300-ms condition, the signal was added to only the end segment, and for all three observers the weight patterns are different from that obtained in experiment 1 with the signal added to all segments. For the 15-ms conditions, altering the signal position changed the estimated weights for only one observer.

Discrimination of narrow-band spectra. I: Spectral weights and pitch cues

The ability to detect changes in the spectral shape of narrow-band tonal complexes (spectral profiles) is examined. The standard consists of three tones of equal intensity; thresholds for detecting an increment in the level of the central, 1000-Hz tone are estimated. A roving-level procedure is used in order to impose a statistical limit on thresholds that can be obtained by basing discriminations on absolute intensity. Estimated thresholds are consistently below this limit, thus indicating the use of other cues. Generally, thresholds are constant for bandwidths ranging from less than a critical band to greater than several octaves. Spectral weight estimates, however, are highly dependent on bandwidth, providing evidence that the discrimination of narrow-band spectra involves different auditory processes than those used to discriminate wideband spectra. Additional data show that pitch cues are important within a restricted range of intermediate bandwidth, but not for wideband or very narrow-band spectra. A version of the EWAIF model involving off-frequency listening is proposed to account for the results.

A temporal model of level-invariant, tone-in-noise detection.

Level-invariant detection refers to findings that thresholds in tone-in-noise detection are unaffected by roving-level procedures that degrade energy cues. Such data are inconsistent with ideas that detection is based on the energy passed by an auditory filter. A hypothesis that detection is based on a level-invariant temporal cue is advanced. Simulations of a leaky-integrator model, consisting of a bandpass filter, half-wave rectification, and a lowpass filter, account for thresholds in band-widening experiments. The decision variable is calculated from the discrete Fourier transform of the leaky-integrator output. A counterintuitive finding is the apparent disassociation of the phenomenon of critical bands estimated from band-widening experiments and the theory of auditory filters. Physiological plausibility is demonstrated by showing that a leaky integrator describes the discharge cadence of primary afferents for tone-in-noise stimuli as well as for complex periodic sounds.

Multiple cues for the discrimination of narrow‐band sounds

The ability to discriminate spectral changes in narrow‐band sounds on the basis of multiple cues is examined. The task is to discriminate a sound consisting of three equal‐intensity tones with frequencies of 960, 1000, and 1040 Hz from a sound with an increment in the level of the center component. In different conditions, listeners are instructed to discriminate the stimuli according to either ‘‘loudness,’’ ‘‘pitch,’’ or ‘‘roughness’’ cues. Spectral weights are estimated in each condition by the method of COSS analysis [B. G. Berg, J. Acoust. Soc. Am. 86, 1743–1746 (1989)]. Relative to a condition without instructions, systematic changes in the pattern of weight estimates are found for each condition. Weight estimates for the loudness and pitch conditions follow patterns predicted by respective models that base discriminations on overall level and on the envelope‐weighted average of instantaneous frequency [L. L. Feth and H. O’Malley, J. Acoust. Soc. Am. 62, 940–947 (1977)] with off‐frequency listening. ...

Effects of randomizing phase on the discrimination between amplitude-modulated and quasi-frequency-modulated tones.

This study investigated the bandwidth of phase sensitivity. Subjects discriminated amplitude-modulated tones (AM), and quasi-frequency-modulated tones (QFM) in a two-interval, forced-choice task. An adaptive threshold procedure was used to estimate the modulation depth needed to discriminate the stimuli as a function of carrier and modulation frequency. Non-monotonicities in threshold-bandwidth functions were often observed at higher modulation frequencies. The results are discussed in terms of two potential cues: (1) waveform envelope, (2) cubic distortion products. In order to degrade the information obtained from auditory distortions, the phase for the carrier frequency was randomly sampled from a uniform distribution, which diminished the non-monotonicities with minimal effect at lower modulation frequencies. Model simulations demonstrated that phase randomization degrades distortion product cues with only a modest effect on temporal cues. Final results show that maximum bandwidths for phase sensitivity (BW(max)) were not proportional to carrier frequencies.