Perceptual systems adjust rapidly to changes in the environment, with neural and behavioral responses dynamically changing to mirror changes in the input (Sharpee et al., 2006; Gutinsky and Dragoi, 2008). Such adaptive codes provide efficient representations because they direct computational resources toward uncommon inputs and provide information about potentially important changes in the world (Barlow, 1990).

Although adaptive coding is less-well-studied in audition than vision, behavioral demonstrations of context-dependence, particularly in speech perception, resonate with the perspective that context adaptively tunes perceptual codes. The identity (Ladefoged and Broadbent, 1957) or accent (Evans and Iverson, 2004) of a talker, the rate of the utterance (Liberman et al., 1956), and the phonetic make-up of a preceding utterance (Mann, 1986) all influence perception of subsequent speech targets. From an adaptive coding perspective, the perceptual system’s response to preceding speech lingers to affect subsequent processing.

Significantly, however, this can only be true to the extent that context and target share common neural resources. In this way, perceptual aftereffects of context have been described as “the psychologist’s microelectrode” (Frisby, 1980) because they can expose neural coding related to perceptual experience through the residual effect of one stimulus upon perception of another (Clifford and Rhodes, 2005). Thus, context-dependent speech perception perhaps can serve to reveal the underlying representation of speech.

A simple model clarifies the approach (Figure 1, after Rhodes et al., 2005). Imagine two populations of units coding a dimension of auditory representational space (e.g., acoustic frequency, or a higher-order feature like talker identity). Pool 1 units best code below-average values along the dimension whereas Pool 2 units better code above-average values; within each pool, more extreme values are coded more robustly. The average value along the dimension is encoded implicitly in the neutral cross-over point at which pools respond equivalently. An input with a higher value along the dimension would result in strong Pool 2 response, thus reducing Pool 2 responsiveness in the short-term due to adaptation. In this way, the context stimulus “lingers” in the perceptual system to affect the resources available to process subsequent stimuli. As is evident in Figure 1, this reduction shifts the neutral point where the two pools respond equivalently toward higher values and the formerly “average” value is now more robustly coded by Pool 1 neurons. Overall, adaptive encoding results in a contrastive shift in target encoding as a function of whether the context was better-coded by Pool 1 or Pool 2 neurons.

This toy model is simple, to be sure, but the general principle may extend to higher-dimensional perceptual spaces and more complex representations. In vision, adaptive coding has been successful in predicting patterns of interactions among low-level representations for brightness and hue (see Frisby, 1980) as well as higher-level representations for faces and bodies (see Clifford and Rhodes, 2005). What is implicit in this approach is the assumption that context and target stimuli are encoded along common dimension(s) of representation. Here, we pursue context-dependent speech categorization to examine significant representational dimensions of speech categories.

For these purposes, Ladefoged and Broadbent’s (1957) classic demonstration of talker normalization is relevant. In their study, listeners heard a constant target-word with a relatively ambiguous vowel at the end of a context phrase “Please say what this word is….” The first (F1) and/or second (F2) formant frequencies (peaks in energy of a voice spectrum; Fant, 1960) of the context phrase were increased or decreased. These shifts can be conceptualized, respectively, as decreasing and increasing the talker’s vocal tract length and, correspondingly, as a change in talker. When the resulting phrases preceded the speech targets, listeners’ categorization shifted in a manner suggesting that they were compensating, or normalizing, for the change in vocal tract length or talker. A constant vowel was more often heard as “bit” when it followed a phrase synthesized as though spoken by a shorter vocal tract (higher formant frequencies in the phrase), but more often as “bet” following the same phrase modeling speech from a longer vocal tract (lower frequencies).

A central and enduring theoretical issue has been the representational dimension across which listeners “normalize” speech categorization in this way. Is the relevant representational dimension talker identity, vocal tract shape/anatomy, or acoustic phonetic space (Joos, 1948; Ladefoged and Broadbent, 1957; Halle and Stevens, 1962; Nordstrom and Lindblom, 1975; McGowan, 1997; McGowan and Cushing, 1999; Poeppel et al., 2008)?

Here, we investigate the extent to which Ladefoged and Broadbent’s classic results can be explained by adaptive coding along a representational dimension that has a general auditory, rather that talker-, or speech-specific basis: the long-term average spectrum (LTAS) of the preceding sound. Recent research has suggested that listeners are sensitive to the LTAS of sound stimuli and adjust perception of subsequent sounds contrastively opposing context LTAS. Holt (2005, 2006a,b) found that sequences of 21 non-speech sine-wave tones, each with a unique frequency sampling a 1000 Hz range affect speech categorization of /ga/–/da/ as a function of the mean frequency of the tones forming the sequence. The influence of these contexts on speech categorization cannot be attributed to any particular acoustic segment of the sequences because tones were randomly ordered on a trial-by-trial basis. Instead, the pattern of context-dependent speech categorization is predicted only by the tone sequences’ LTAS. Perception of the subsequent speech targets was relative to, and contrastive with, the LTAS consistent with the adaptive coding scheme sketched above if LTAS serves as the common representational dimension linking non-speech contexts and speech targets.

Here, we seek to directly replicate the Ladefoged and Broadbent (1957) results with speech contexts and to explicitly test whether non-speech contexts modeling critical characteristics of the speech contexts’ LTAS produce the same effects on vowel categorization. Directly mimicking the methods of Ladefoged and Broadbent (1957), listeners categorized a series of speech sounds varying perceptually from “bet” to “but” in the context of a preceding phrase (“Please say what this word is…”). This phrase was synthesized to model two different “talkers”: one with a larger vocal tract and the other with a smaller vocal tract. The same listeners also categorized the same “bet” to “but” targets preceded by sequences of non-speech sine-wave tones with frequencies modeling the LTAS of the context phrases. These extremely simple acoustic contexts carried no talker- or speech-specific information. Should the speech and non-speech contexts similarly influence vowel categorization, it would suggest that speech and non-speech contexts draw upon common neural resources. Rather than articulatory or talker-specific dimensions that typically have been proposed to account for talker normalization, listeners may rely on sounds’ LTAS to tune speech perception.

Participants’ categorization of speech targets in isolation was orderly, as indicated by a significant main effect target F2 frequency, F (5, 25) = 139.43, p < 0.01. Individuals’ data conformed to this average pattern.

Figures 2B,C illustrate the influence of context on vowel categorization. A 2 (context type, speech/non-speech) X 2 (LTAS, low/high) X 6 (target F2 frequency) repeated-measures ANOVA of listeners’ percent “but” responses reveals that, as expected from vowel categorization in isolation, responses varied reliably as a function of the target F2 frequency, F (5, 25) = 290.50, p < 0.001. Moreover, there was no main effect of context type indicating no overall bias in vowel categorizations a function of the type of context that preceded targets, F (1, 25) = 0.274, p = 0.61.

Of greater interest, there was a significant main effect of LTAS on vowel categorization, F (1, 25) = 27.05, p < 0.001. The vowels were categorized as /∧/ significantly more often following high-frequency contexts whereas the same vowels were more often labeled as /ε/ following low-frequency contexts. Thus, the context-dependent effect was spectrally contrastive and consistent with previous studies of the influence of sentence-length contexts on speech categorization (Ladefoged and Broadbent, 1957; Watkins and Makin, 1994, 1996; Holt, 2005, 2006a,b; Huang and Holt, 2009). The interaction between LTAS and target F2 frequency was significant, F (5, 25) = 11.65, p < 0.001, indicating that context had a greater influence on perceptually ambiguous targets.

This pattern of contrastive context-dependent vowel categorization was evident for both speech, F (1, 25) = 21.62, p < 0.001, and non-speech contexts, F (1, 25) = 17.38, p < 0.001. Of primary interest, there was no significant interaction between context type and LTAS, F (1, 25) = 0.324, p = 0.574. It is interesting that although the LTAS contrast was larger in the non-speech contexts compared with speech context condition (Figure 3; see detail explanation in discussion), the magnitude of the influence of speech and non-speech contexts on speech target categorization was statistically indistinguishable. Neither the interaction between context type and target frequency, F (1, 25) = 1.22, p = 0.30, nor the three-way interaction was significant, F (5, 25) = 1.424, p = 0.22. In sum, “talker” is not an essential element of talker normalization as it appears even in the absence of a talker when context is merely a sequence of sine-wave tones.

We exploited context-dependent speech categorization as “the psychologist’s microelectrode” (Frisby, 1980) to reveal characteristics of the underlying representation of speech. The residual effect of one stimulus upon perception of another demands that the two share common neural processing and/or representation. Thus, the comparable influence of speech and non-speech contexts on speech categorization indicates a common substrate of interaction. Importantly, since the two context types did not share linguistic, articulatory gestural, or talker-specific information, but yet produced equivalent effects on speech categorization, it does not appear that speech-, vocal tract-, or talker-specific information is essential in eliciting the patterns of context-dependent perception that have been described in the literature as “talker normalization.”

What the two context types in the present experiment shared was a similar pattern of spectral energy across their time course. Figure 3 illustrates the LTAS of the speech-target endpoints (3A) and speech (3B) and non-speech contexts (3C). Gray shading highlights the region of acoustic energy that critically distinguishes the speech targets. We suggest that the context-dependent speech categorization observed here (and in Ladefoged and Broadbent, 1957) arises because the auditory system is sensitive to the context LTAS and the speech target is encoded relative to, and contrastive with, that long-term average.

Specifically, we propose that speech is adaptively coded according to the LTAS of ambient sound. The simple model described in the introduction can clarify one means by which this might be accomplished. Imagine pools of neurons sensitive to energy in the range of the second formant with one pool better coding relatively lower frequencies and the other better coding higher frequencies. By this model, presentation of the speech context modeling a longer vocal tract with lower-frequency energy within this frequency range would result in greater activity among the pool of neurons that better code lower frequencies. Subsequent adaptation among this pool of responsive neurons would result in a shift toward the opposite, higher-frequency neural pool at the time of speech target presentation. The formerly “neutral” frequencies would be now more robustly encoded by the higher-frequency neural pool, shifting representation contrastively away from the lower-frequency context. This adaptive coding serves to exaggerate differences between the LTAS of the context and target (Holt, 2006a).

By this model, “talker normalization” effects on speech targets are predicted and obtained even when no speech information is available in the context. Of note, the LTAS model makes no reference to specific linguistic units, such as phonemes. This generality makes the adaptive coding approach in general, and the LTAS model in particular, extend straightforwardly to other normalization phenomena in speech perception. Huang and Holt (2009) report that shifts in the peak energy of the LTAS in the region of the fundamental frequency (f0) of a Mandarin Chinese sentence predict patterns of context-dependent Mandarin lexical tone normalization (Leather, 1983; Fox and Qi, 1990; Moore and Jongman, 1997). Further, non-speech precursors with matched LTAS produce the same effects. Similarly, although the adaptive coding approach to context-dependent speech categorization reveals LTAS as an important dimension of representation, the model’s application is more general. Wade and Holt (2005), for example, investigated rate-dependent normalization effects whereby the rate of a precursor sentence affects how listeners categorize a rate-dependent speech distinction like /ba/ versus /wa/, finding that the rate of presentation of a sequence of non-speech tones evokes a similar contrastive influence on speech categorization.

If demonstrations of “talker normalization” can be accounted for by general perceptual processes that are not talker- or speech-specific then there remains the question of whether the context-dependent speech categorization taken as evidence of normalization really accommodates the acoustic variability in speech that arises from different talkers. We argue that it does. Across context sentences like those of the current study, speech maps the scope of a talker’s articulatory space and the mean of that space resembles a talker’s neutral vowel, the shape of the non-articulating vocal tract (Story, 2005). This neutral vowel serves as an effective normalization referent because, as Story (2005) has demonstrated, most of the variability across talkers can be accounted for by differences in the shape the vocal air space of talkers’ neutral vowels. Thus, if listeners were able to extract an estimate of the neutral vowel, the mean of the articulatory space, they would have an excellent referent for talker normalization. However, tracking back from acoustics to articulator requires negotiating the inverse problem. The results we describe here suggest an alternative.

As talkers produce a variety of consonants and vowels, speech maps the articulatory space but it also produces a sound spectrum that maps the acoustic space and, across time, samples an LTAS. Instead of solving the inverse problem to recover the actual neutral vocal tract shape as an articulatory referent for normalization, listeners may use the average spectrum LTAS as an auditory referent. The observation that non-speech tones modeling the LTAS of a talker are as effective in shifting speech categorization as speech contexts supports the viability of a general auditory referent. However, it should be noted that LTAS is unlikely to be the only contributing factor in talker normalization. Speaker identities, for example, may mediate listeners’ attention and expectation and influence the way listeners tune their perception to the preceding contexts (Magnuson and Nusbaum, 2007). Talker normalization is likely to be a multi-facet phenomenon. Nonetheless LTAS, which provides sufficient context information via general perceptual processes, is an important factor in the adaptive coding of speech perception processing that contributes to talker normalization.

The lingering influence of a sequence of non-speech tones on listeners’ response to speech targets indicates a shared substrate, the level of which remains to be evaluated. Relevant to this, Holt (2005) reported that non-speech contexts influence on speech categorization persists across 1300 ms of silence. Moreover, the aftereffect of the tones was present even when 13 neutral-frequency tones intervened between tone contexts and speech targets. The influence of temporally non-adjacent tone sequences can even override the influence of temporally adjacent speech contexts on speech targets (Holt, 2006b). These observations argue for a central, rather than peripheral, substrate.

In fitting the mind to the world, the ambient context plays a large role in how input is coded. The present results demonstrate that patterns of context-dependent speech categorization long taken to be evidence of talker-specific normalization for articulatory referents or rescaling of phonetic space may arise, instead, from general principles of adaptive perceptual coding. In “listening for the norm,” listeners appear to adjust perception to regularities of the ambient environment.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.