One of the most basic questions of cognitive science is how do organisms sort the objects of the world into categories? Categorization is essential to all cognitive processes. Without categorization, each object would be perceived as unique and no generalization rules could be used to take rapid and appropriate decisions [1]. However, because no single perceptual feature is likely to be a necessary and sufficient condition for category membership [2], identifying the neural substrates underlying categorization processes is a real challenge.

One example of categorical perception is the perception of phonemes in human speech [3]. However, categorical perception is specific neither to humans nor to speech. Indeed, chinchillas have been shown to categorically perceive speech much the same way as humans do [4]. There has also been evidence of categorical perception of species-specific vocalizations in monkeys [e.g. 5] and in avian species [e.g. 6].

Among birds, songbirds are particularly interesting because, without necessarily requiring categorical perception per se, birdsongs often provide researchers with clear examples of categories of signals allowing species, population and individual recognition. For example, swamp sparrows sing two note categories with different roles in song construction, and there has been evidence of categorical perception for these notes by this species [7]. Budgerigars, which, like songbirds, are vocal learners, have been shown to group vocal stimuli according to functional and acoustical categories [6].

These birds also have the advantage of possessing a system of specialized brain structures found only in birds that learn to sing: the song system [8]. Given the many parallels that exist between birdsong learning and speech development [9]–[11], this system has become a choice model for studying the neural bases of vocal communication [12]. Nuclei involved in song production have thus been well characterized [e.g. 13] but less is known about areas involved in song perception and discrimination [for a recent review, see 14]. Recently however, the caudo-medial telencephalon has emerged as a plausible site for sensory representation of birdsong.

Caudo-medial telencephalon contains thalamo-recipient Field L2, which is comparable to thalamo-recipient layer IV of the mammalian auditory cortex, and two of its targets, caudo-medial nidopallium (NCM) and caudo-medial mesopallium (CMM), which can be compared to supragranular cortical layers [15]. Based on electrophysiological responses [16]–[19] or on the expression of an immediate early gene (IEG–ZENK) [20]–[22], NCM auditory responses have been shown to be the strongest for conspecific songs, followed by heterospecific songs and non-song acoustic signals, and they are known to show a rapid and long-lasting habituation effect that is song-specific. NCM neurons thus appear to be able to discriminate between different conspecific songs, a property that is required for perceptual song discrimination. Moreover, the ZENK response to a social stimulus has been observed to be proportional to the animal's preference for this stimulus: for example in female European starlings, where a preference for male long bout songs over male short-bout songs has been observed, NCM appears to show higher expression of ZENK in response to long-bout than to short-bout songs, independently of the total amount of song that the females heard [23]. It has therefore been suggested that NCM may serve as a common source for behaviourally relevant distinction among conspecific song features, which would be then extracted by different higher processing areas [see also 24]–[27].

Previous studies thus point to NCM as a well positioned brain region for categorization of songs with different functions and social values. For this reason, we hypothesized that, in our model songbird, which is the starling, NCM neurons may respond differentially to the distinct functional classes of songs that have been described in this highly social species. Indeed, male starlings sing three structurally and functionally distinct classes of songs that are used for species, population and individual recognition (see Fig. 1 for examples) [28]. Class-I songs are short, simple and loud whistles sung by all male starlings that are used in species and population recognition (dialectal variants) [29], [30]. Class-II songs are also short, simple and loud whistles that are used in individual recognition, especially between same-sex social partners [31]. Finally, class-III songs, also called warbling, are long, complex and soft songs that are used in individual recognition at short distance, especially between males and females [32], [33]. These three classes of songs differ not only by their structure [34], [35], but also by their pattern of acquisition during song learning [36]–[38], and by their context of emission [29], [39]–[42]. They thus correspond to clear and distinct classes of sounds with different functions and social values. Hence, we used these songs to test for potential differential responses in the NCM of awake-restrained adult male starlings and for a possible correspondence between these responses and the functional aspects of songs. Our study demonstrates that the activity of NCM neurons can indeed indicate or represent a class of sounds corresponding to a behaviourally-defined recognition process.

Using a multi-electrode array and a mapping method based on systematic recordings (no search stimulus; see Material and Methods), we recorded the electrophysiological activity of 1972 neuronal sites in the NCM of 6 awake-restrained adult male starlings (mean±SE = 328.7±21.4 sites/bird; 188.2±26.9 sites in the left hemisphere, 140.5±18.3 sites in the right hemisphere; Wilcoxon, p = 0.07), while broadcasting artificial non-specific sounds (pure tones and white noise) and natural species-specific stimuli corresponding to the three classes of songs described in starlings (see Fig. 1). Among these sites, 32.3% were responsive to at least one of the stimuli we used (mean±SE = 31.4±7.4% in the left hemisphere, 35.6±6.5% in the right hemisphere; Wilcoxon, p = 0.25). Only these responsive sites (n = 633; mean±SE = 105.5±22.2 responsive sites/bird) were further analyzed.

The relative frequency distribution of sites responding to 1, 2, 3,… or 23 (that is all) stimuli is shown in Figure 2. On average, responsive sites responded to 6.5±0.6 stimuli (mean±SE = 6.5±0.8 stimuli in the left hemisphere, 6.5±0.6 in the right hemisphere; Wilcoxon, p = 0.34), that is about 30% of the stimuli used. About 30% of the responsive sites responded to only 1 or 2 stimuli, and no site responded to all stimuli. Overall, more than 70% of the sites that responded to only one stimulus responded to class-II or –III stimuli, and less than 10% responded only to nonspecific stimuli (note that we could observe no site responding only to the 4- and 8-kHz pure tones and to the white noise).

When responses to each stimulus were considered, it appeared that, globally, many more sites responded to natural species-specific sounds than to artificial non-specific stimuli (with overall 4.2–15.7% of the responsive sites responding to artificial non-specific stimuli vs. 8.3–56.9% responding to natural species-specific sounds; see Fig. 3). However, the different species-specific song types were not equally effective at driving neuronal responses, and some of them even elicited responses at a level similar to some artificial non-specific sounds, such as some of the class-I songs (the simple themes, with 17.5 and 18.0% of the responsive sites responding; Fig. 3) and the species-specific class-III motif (clicks, with 8.3% of the responsive sites responding; Fig. 3). In fact, most sites responded to songs used in individual recognition, such as the class-II whistles (with 41.0–48.4% of the responsive sites responding; Fig. 3) and the individual-specific class-III motifs (with 56.1 and 56.9% of the responsive sites responding; Fig. 3).

Most importantly, when we considered responses to each class of stimuli, it appeared that both the proportion of responsive neuronal sites (Fig. 4) and the magnitude of the neuronal responses (as measured by Z scores; Fig. 5) significantly differed from one class to another: responses were the strongest for the highly individual class-III motifs, followed by the individual-specific class-II songs, the species-specific class-I whistles, and finally the artificial non-specific stimuli (two-way repeated-measures ANOVAs and PLSD Fisher tests, stimulus class effect: p<0.0001 for both the proportion of responding sites and the Z scores, post-hoc comparisons: p<0.05 for all pairwise comparisons in both cases, no hemisphere effect, no interaction). These differences were neither due to a specific bird (since within-bird comparisons showed the same effect or trend in each bird) nor to one particular subset of stimuli (as, with the exception of the species-specific clicks in class III, proportions of sites responding to each stimulus appeared to be relatively homogenous within each class; see Fig. 3). Thus, intra-class variations appeared to be lower than inter-class variations, especially for class-II and individual-specific class-III stimuli which showed coefficients of variation (CVs) that were 3 to more than 7 times lower than the CV observed across all stimuli (mean CVs for class II = 21 and 22%, for class III w/o clicks = 12 and 9% and for all stimuli = 64 and 69%, respectively for the left and right hemispheres).

Finally, for classes within which stimuli were familiar or not (classes II and III, see Fig. 1), no effect of familiarity could be observed: comparisons of unfamiliar, familiar and bird's own songs within classes II and III showed no significant difference between these stimuli, neither in the proportion of responding neuronal sites (see Fig. 3; two-way repeated-measures ANOVAs, p = 0.26 and 0.71 respectively for classes II and III) nor in the magnitude of the neuronal response (two-way repeated-measures ANOVAs, p = 0.21 and 0.60 respectively for classes II and III).

Our results show differential neuronal responses to ethologically defined classes of songs that have different functions and social values, in a non-primary, associative auditory area, both in the number of neuronal sites responding, and in the magnitude of the responses. Most importantly, these differential responses corresponded to the functional classes of songs, with increasing activation from non-specific to species-specific and from species-specific to individual-specific sounds. Our data thus provide a rare example of convergence between natural behaviour and neural activity, and point to a potential neural substrate for categorization of complex communication signals and vocal recognition of same-species members.

These results suggest that NCM could be the place for sorting sounds into categories in the songbird brain. Thus, although we did not (and could not) test a continuum of sounds from one category to another, the proportion of responding sites and the magnitude of responses clearly differed between functional classes of songs and less within them. Given the large structural variations within each class (e.g. between different individual-specific class-II whistles or warbling motifs; see Fig. 1), it is therefore very likely that NCM differential responses relied more on the stimuli behavioural salience than on their acoustic structure. Another element in this direction is the finding that clicks, which, contrarily to other warbling motifs, are common to all starlings, and which are normally produced within a sequence of other motifs, were processed similarly to the very dissimilar structures that are pure tones or white noise. Finally, the fact that familiarity did not modify within-class responses demonstrates a capacity of generalization, which is one of the cue properties of categorization.

Several authors have questioned the role of the NCM in the emergence of the highly complex and song-selective responses observed in higher-order song nuclei [16]–[22]. Here, we show that NCM neurons appear to differentially respond to distinct classes of songs corresponding to different levels of recognition (going from species-specific to individual-specific). Interestingly, the most effective stimuli at driving the neurons were the highly individual class-III motifs, which are thought to be the basic units of individual recognition in starlings [43], [44]. NCM neurons could therefore actually contribute to the processing resulting in the strong selectivity for both spectral and temporal properties of song observed in higher-order regions of the avian brain [45]–[47]. This selectivity is known to arise during development in neurons that are initially unselective [48]. One can therefore wonder what happens at the level of the NCM, and how the differential responses we observed develop. Evidence exists that NCM neuronal responses are shaped by the animal's prior experience with song [49]. Moreover, we have shown experience-dependent effects at the level of the main input to NCM (the Field L, which is comparable to the mammalian primary auditory cortex) [50], [51]. Given that, in a hierarchical scheme of sensory processing, plasticity at the primary levels should influence higher-order regions, it is very likely that NCM response properties are also expressly determined by the animal's unique experience. This will have to be confirmed by studying how these response properties develop, for example in birds lacking experience with one or several classes of songs for which we here observed differential neuronal responses. We already know that, in young starlings' vocalizations, some classes of songs can be absent according to the birds' experience [36], [38]. We also know that class-I songs are learned later in development [40]. Interestingly, this class of songs was the least effective natural class of songs in our study. We could therefore imagine that NCM responses may modulate input to the song system, thus playing the role of an “intrinsic perceptive filter” [52] and favouring learning of class-III songs first, class-II songs next, and finally class-I songs. Such a mechanism would be somewhat reminiscent of what is thought to happen in human infants, who may use the acquired phonological categories of the language to which they are exposed to guide speech development [53].

The strong similarities that the process of vocal imitation through which songbirds learn their vocalizations presents with the mechanisms underlying the ontogeny of human language have been known for long [9]–[11], and they have made songbirds the most relevant biological model to understand language acquisition and its neural correlates [54], [55]. Much more recently, a consortium of avian brain specialists has also unveiled similarities in the brain structures of birds and mammals [56], [57]. Some authors have thus compared NCM to the superficial layers of auditory cortex or to secondary mammalian auditory regions such as the lateral belt in primates or Wernicke's area in humans [14], [58]. We believe our study reinforces the parallel between these structures, and widen the impact of studies on songbirds. Indeed, we here observed differential representation of sounds with distinct biological significance (as observed in the field), as seen in higher-order fields of the auditory cortex of mammals, including humans. For example, differences in the biological significance of calls, as expressed in behavioural tests, can be seen in the activation of the secondary auditory field of mice [59]. Neurons in higher-order auditory cortical fields of monkeys prefer complex spectro-temporal acoustic patterns compared to simple sounds [60]. Finally, functional imaging studies in humans have shown a stronger activation in response to speech than in response to non-speech sounds in areas outside the primary auditory cortex in the temporal lobe [61]. Our results therefore show striking parallels with what has been observed in mammals, including humans, and point to an associative auditory area that provides starlings, and probably other higher vertebrates, with a potential neural substrate for the extraction of biologically relevant information contained in complex acoustic signals used in vocal communication. Overall, by indicating that NCM as a brain area is a likely candidate for sound categorization in birds, we believe that the present study improves our knowledge of the representation of sound significance in the songbird brain, and that it will participate in comparing sound categorization in vertebrates.