Although the relations between second formant (F2) onset and F2 vowel are extremely regular and contain important information about place of articulation of the voiced stops, they are not sufficient for its identification. Using quadratic discriminant analysis of a new data set, it is shown that F3 onset and F3 vowel can also contribute substantial additional information to help identify the consonants.

The idea of the orderly output constraint is compared with recent findings about the representation of vowels in the auditory cortex of an animal model for human speech sound processing (Ohl & Scheich 1997). The comparison allows a critical consideration of the idea of neuronal “feature extractors,” which is of relevance to the noninvariance problem in speech perception.

Evaluating the current locus equation under ideal conditions identifies important and unexpected parameter dependencies. Locus equation (LE) utility, either as a valid laboratory tool or possible invariant cue, depends on stringent specification of critical parameters and rigorous empirical testing.

Neurobiologically inspired theories of speech perception such as that proposed by Sussman et al. are useful to the extent that they are able to constrain such theories. If they are simply intended as suggestive analogies, their usefulness is questionable. In such cases it is better to stick with the Gibsonian approach of attempting to isolate invariants in speech and to demonstrate their role for the perceiver in perceptual experiments.

The arguments for the orderly output constraint concern phylogenetic matters and do not address the ontogeny of combination-specific neurons and the corresponding processing mechanisms. Locus equations are too variable to be strongly predetermined and too inconsistent to be easily learned. Findings on the development of speech perception and underlying auditory processing must be taken into account in the formulation of neural encoding theories.

This commentary questions the proposed resemblance between the auditory mechanisms of localization and those of the sensory registration of speech sounds. Comparative evidence, which would show that the neurophysiology of localization is adequate to the task of categorizing consonants, does not exist. In addition, Sussman et al. do not offer sensory or perceptual evidence to confirm the presence in humans of processes promoting phoneme categorization that are analogous to the neurophysiology of localization. Furthermore, the computational simulation of the linear model of second formant variation is not a plausible sensory mechanism for perceiving speech sounds.

We consider Sussman et al.'s suggestion that auditory biases for processing low-noise relationships among pairs of acoustic variables is a preadaptation for human speech processing. Data from other animal communication systems, especially those involving sexual selection, also suggest that neural biases in the receiver system can generate strong selection on the form of communication signals.

A tendency of auditory cortical neurons to respond at the beginning of major transitions in sounds rather than providing a continuously updated spectral-temporal profile may impede the generation of combination-sensitivity for certain classes of stimuli. Potential consequences of the cortical encoding of voiced stop-consonants on representational principles derived from orderly output constraints are discussed.

Locus equations are supposed to capture a perceptual invariant of place of articulation in consonants. Synthetic speech data show that human classification deviates systematically from the predictions of locus equations. The few studies that have contrasted predictions from competing theories yield mixed results, indicating that no current theory adequately characterizes the perceptual mapping from sound to phonetic symbol.

The potential role of locus equations in three existing models of human classification behavior is examined. Locus equations can play a useful role in single-prototype and boundary-based models for human consonant recognition by reducing model complexity.

Neural algorithms are conserved during evolution. Neurons with different shapes and using different molecular mechanisms can perform the same computation. However, evolutionary conservation of neural algorithms is not sufficient for claiming the realization of an algorithm for a specific computational problem. A plausible scheme for ontogenetic emergence of the structure of the algorithm must also be provided.

Combination-sensitive neurons serve as the fundamental processing unit in Sussman and colleagues' proposal for the neural representation of stop consonants. This commentary describes recent studies in the mustached bat that show how ubiquitous and flexible this neural strategy can be. Sussman et al.'s proposal is an important contribution to a neuroethological consideration of speech perception.

The most frequent criticism of the target article is the lack of clear separability of human speech data relative to neuroethological data. A rationalization for this difference was sought in the tinkered nature of such new adaptations as human speech. Basic theoretical premises were defended, and new data were presented to support a claim that speakers maintain a low-noise relationship between F2 transition onset and offset frequencies for stops in pre-vocalic positions through articulatory choices. It remains a viable and testable hypothesis that the phenomenon described by the locus equation is a functional adaptation of production mechanisms to processing preferences of the auditory system.

What R&P term the “noise reduction model” implies that the psi-conducive state is related to the induction of an altered state of consciousness (ASC). Yet there is a problem in embedding psi in the ASC, because one anomaly is replacing another. This seems to be a general strategy in the literature of the anomalous.

Pseudoscience is not an appropriate label for parapsychology. Although the noise reduction model of extrasensory perception (ESP) is explanatory only in a limited sense, research does exist addressing the correlation between ESP and altered states of consciousness (ASCs). The term anomaly is not appropriately applied to experiences such as out of body experiences (OBEs) per se, but only to the question of their source. Research on both topics should be encouraged.

My criticisms of parapsychology are neither based on its subject matter per se, nor simply on a charge of sloppy research, but rather on the whole pattern of theory and research in this domain. The lack of a positive definition of psi, the use of ad hoc principles such as psi-missing and the experimenter psi effect to account for failures to confirm hypotheses, and the failure to produce a single phenomenon that can be replicated by neutral investigators are among the major problems that keep parapsychology outside regular science. Glicksohn and I agree that anomalous experiences should be investigated.

Wilson and Sober's (1994t) revival of group selection theory may have failed with some readers because its simple arithmetic foundation was obscured under the complexities of its presentation. When that uncontrovertible principle is uncovered, it broadens dramatically the fundamental motives that social scientists may impute to human nature and still be consistent with Darwinian evolutionary theory.

We reinforce Thompson's points by providing a second example of the paradox that makes group selection appear counterintuitive and by discussing the wider implications of multilevel selection theory.

Clark & Thornton take issue with my claim that parity is not a generalisation problem, and that nothing can be inferred about back-propagation in particular, or learning in general, from failures of parity generalisation. They advance arguments to support their contention that generalisation is a relevant issue. In this continuing commentary, I examine generalisation more closely in order to refute these arguments. Different learning algorithms will have different patterns of failure: back-propagation has no special status in this respect. This is not to deny that a particular algorithm might fortuitously happen to produce the “intended” function in an (oxymoronic) parity-generalisation task.

In his new commentary, Damper re-emphasises his claim that parity is not a generalisation problem. But when proper account is taken of the arguments he puts forward, we find that the proposed conclusion is not the only one that can be drawn.