Introduction

A large body of behavioral and neuroscience research suggests that visual category learning is mediated by at least two separate, albeit partially overlapping, learning systems (Ashby & Maddox, 2005; Knowlton, 1999; Nomura & Reber, 2008; Poldrack & Packard, 2003). The explicit, reflective learning system depends on working memory and executive attention to develop and test hypotheses and rules for explicit classification. Processing in this system is available to conscious awareness and is mediated by a circuit primarily involving the dorsolateral prefrontal cortex, anterior cingulate, and anterior caudate nucleus (Ashby & Ell, 2001; Seger & Miller, 2010). The implicit, procedural-based, reflexive learning system is not consciously penetrable and operates by associating perception with actions that lead to reinforcement via feedback. Dual-system models predict that the two systems are complementary in learning various category structures, some of which are reflective-optimal, and others reflexive-optimal. Although more than 20 years of research has motivated the dual-system framework, this model has not been systematically applied to examine speech category learning.

Previous speech-learning studies have examined category learning as an emergent property of unsupervised and/or supervised learning processes (Goudbeek, Cutler, & Smits, 2008; McClelland, Fiez, & McCandliss, 2002; Norris, McQueen, & Cutler, 2003; Toscano & McMurray, 2010; Vallabha, McClelland, Pons, Werker, & Amano, 2007). In unsupervised learning, statistical regularities in the input lead to category representations in sensory regions through a process of implicit Hebbian learning (Goudbeek et al., 2008; Goudbeek, Swingley, & Smits, 2009; McClelland et al., 2002). More recent, computationally based, unsupervised-learning models incorporate competition in addition to statistical learning (McMurray, Aslin, & Toscano, 2009; Toscano & McMurray, 2010). From the neurobiological perspective, unsupervised category learning is instantiated within topographical maps in the primary and secondary auditory regions that are sensitive to input statistics (Guenther, Nieto-Castanon, Ghosh, & Tourville, 2004; Vallabha & McClelland, 2007). In contrast, supervised-learning models posit that some form of instructional feedback (lexical or selective attention) to the sensory network is necessary, in addition to Hebbian learning (Norris et al., 2003). Although significant unsupervised speech learning can occur in adults, category learning with feedback can lead to substantially larger gains (Goudbeek et al., 2008; McClelland et al., 2002).

While the role of statistical learning and feedback instruction in the form of lexical influences has been extensively researched in speech learning, there has been less focus on the role of domain-general feedback-based learning processes in mediating category-learning success. This is despite the fact that functional neuroimaging studies examining speech category learning in adults implicate reflective and reflexive learning circuitry in addition to the auditory regions (Callan et al., 2003; Tricomi, Delgado, McCandliss, McClelland, & Fiez, 2006). In dual-system models, reflective rules are encoded within the sensory areas with bidirectional connections to working memory units within the lateral portion of the prefrontal cortex (PFC). When a new rule is generated, the excitatory input from the PFC to the head of the caudate is strengthened, resulting in the maintenance of a newly established rule. The PFC units, each of which represents a particular rule, are activated by the anterior cingulate to select among various alternative rules. In comparison, during reflexive learning, a single striatal unit (or small group of units) implicitly associates an abstract cortical–motor response with a large group of sensory cells. The critical aspect of learning occurs at cortical–striatal synapses, and synaptic plasticity is facilitated by a dopamine-mediated reinforcement training signal. Despite the different circuitries, both the reflective and reflexive learning systems utilize the sensory component within the primary and association auditory regions. These components are reflectively or reflexively associated with rewards (e.g., instructional feedback).

Our working hypothesis is that speech categories are optimally learned by the reflexive learning system. This is because speech categories are often difficult to verbalize, are easily learned by infants whose attention and working memory networks are immature (Echols, 1993; Mugitani et al., 2009; Pierrehumbert, 2003), and utilize acoustic cues that are multidimensional, highly redundant, and variable across talkers (Gandour, 1983; Holt & Lotto, 2008, 2010). Creating rules for such a large dimensional space may not be optimal, since generating and testing rules that involve multiple dimensions is resource intensive. In the present article, we use training manipulations on trial-by-trial feedback (Experiments 1 and 2) and talker variability (Experiment 3) to examine the relative contribution of the reflective and reflexive learning systems to speech learning success.

The reflective and reflexive learning systems respond differentially to various training manipulations. For example, delaying the presentation of feedback impairs learning in the reflexive system, but not in the reflective system (Maddox, Ashby, & Bohil, 2003). This is because the reflexive system is critically dependent on dopamine-mediated stimulus–response implicit reward learning. Delaying feedback interferes with dopamine release, reducing the effectiveness of the association of stimulus–response with reward. Also, “full” feedback that provides the correctness of the response on each trial, as well as information about which category was present, speeds learning in the reflective system (Maddox, Love, Glass, & Filoteo, 2008), relative to “minimal” feedback that provides only the correctness of the response on each trial. Full feedback promotes the generation and testing of rules that are critical to reflective learning but disrupts the transfer of control to the reflexive system (Maddox et al., 2008). Previous studies have used these timing and feedback manipulations to dissociate the learning systems in artificial category learning, but not in natural speech category learning.

To this end, we conducted three category learning experiments to examine the effect of various training manipulations that target either the reflective or the reflexive learning systems on speech category learning success. In each experiment, native English speakers were trained to categorize nonnative Mandarin tone categories produced by multiple talkers (Fig. 1) with instructional feedback. This type of training structure (trial-by-trial feedback, high talker variability) is ubiquitous in the speech learning literature. In Mandarin Chinese, tone contours signify differences in word meaning (e.g., /ma/ with a rising tone means “mother,” while /ma/ with a falling tone means “to scold”; see Fig. 1). Previous studies have shown that native English speakers have difficulty in learning tone categories, which is hypothesized to result from inadequate relative weighting of talker-independent pitch direction cues (Chandrasekaran, Sampath, & Wong, 2010; Wang, Jongman, & Sereno, 2003).

Fig. 1
figure 1

Example of multiple-talker stimuli used in the category training study. Fundamental frequency contours of the four tones (T1 = high-level; T2 = low-rising; T3 = low-dipping; T4 = falling) produced by four native Mandarin speakers (two female). Tone contours were obtained using Praat (Boersma & Weenink, 2011)

Full size image

The training manipulations used in Experiments 1 and 2 were derived from visual category learning studies. Experiment 1 determined the extent to which the immediacy of feedback (immediate vs. delayed) impacts tone category learning. Experiment 2 determined the extent to which information content of feedback (full vs. minimal feedback) impacts tone category learning (Fig. 2). Immediate feedback is critical for the reflexive system but not the reflective system (Maddox et al., 2003), while full feedback selectively speeds reflective learning but impairs reflexive learning (Maddox et al., 2008). On the basis of our working hypothesis, we predicted that feedback manipulations that targeted the reflexive learning system (immediate or minimal feedback) would enhance learning, relative to those that target the reflective learning system (delayed or full feedback).

Fig. 2
figure 2

Experimental procedures. In Experiments 1–3, we examined the effects of reflexive (top) or reflective (bottom) training manipulations on tone category learning success

Full size image

While dual-system models of visual category learning make specific predictions about feedback processing, they offer no clear prediction about the impact of talker variability on category learning success, which is argued to be important for generalization to new talkers (Lively, Logan, & Pisoni, 1993) and association of categories with more reliable acoustic cues (Apfelbaum & McMurray, 2011; Rost & McMurray, 2009). While most agree that multitalker training is advantageous, the role of the order of talker presentation, if any, has not been systematically examined (although see Perrachione, Lee, Ha, & Wong, 2011, which we return to in the Discussion section). Within the framework of the dual learning systems, we predicted that systematically blocked talker presentation would promote reflective learning, whereas a randomly mixed talker presentation would enhance reflexive learning. Our logic here is that blocked talker presentation promotes faster hypothesis testing and validation and is, therefore, less resource intensive for the reflective system than is the mixed talker condition. Indeed, previous neuroimaging work has shown that the mixed talker condition engages the frontal working memory system more extensively than does the blocked talker condition (Wong, Nusbaum, & Small, 2004). Furthermore, the mixed talker presentation does not allow learners to predict the next talker in advance, disrupting the reflective generation and testing of talker-specific rules. Therefore, learners are more likely to associate talker-invariant acoustic cues (e.g., pitch direction) with implicit reward than talker-variant cues (e.g., pitch height). On the basis of the hypothesis that speech learning is optimally learned by the reflexive learning system, we predicted enhanced learning in the mixed talker condition, relative to the blocked talker condition.

To summarize, enhanced learning in the immediate feedback condition relative to delayed feedback (Experiment 1), minimal feedback relative to full feedback (Experiment 2), and mixed talker condition relative to blocked talker condition (Experiment 3) will be considered as support for the dominance of the reflexive system in speech learning during adulthood.

Method

Participants

Undergraduate students at the University of Texas were recruited (n = 194; age range: 18–35 years) and monetarily compensated for their participation. Participants reported no history of neurological or hearing deficits and were native speakers of American English, with no prior exposure to a tone language. Music history questionnaires were collected to match the groups on musicianship (Wong, Perrachione, & Parrish, 2007). All participants provided informed consent and were debriefed following the experiment. In Experiment 1, participants were divided into immediate (n = 25; 15 female) and delayed (n = 30; 14 female) feedback groups (equivalent years of musical training, p = .585; immediate, mean = 3.42, SEM = .675; delayed, mean = 2.90, SEM = .664). In Experiment 2, participants were divided into minimal (n = 41; 20 female) and full (n = 40; 21 female) feedback groups (equivalent years of musical training, p = .979; minimal, mean = 2.20, SEM = .525; full, mean = 2.21, SEM = .393). In Experiment 3, participants were divided into mixed (n = 30; 18 female) and blocked (n =28; 15 female) talker groups (equivalent years of musical training, p = .723; mixed, mean = 3.18, SEM = .787; blocked, mean = 3.57, SEM = .776). Participants did not overlap between groups.

Stimuli for tone category training

Four native Mandarin Chinese speakers (2 female) originally from Beijing produced four Mandarin tones: tone 1 (T1; high-level), tone 2 (T2; low-rising), tone 3 (T3; low-dipping), and tone 4 (T4; high-falling; see Fig. 1). The tones were produced in citation form in the context of five monosyllabic Mandarin Chinese words (bu, di, lu, ma, and mi), reflecting variability inherent in natural language. The 80 stimuli were RMS amplitude and duration normalized (70 dB, 0.44 s). Duration normalization was achieved using the PSOLA (Pitch Synchronous Overlap and Add) module incorporated within the Praat software (Boersma & Weenink, 2011). Five independent native Mandarin Chinese speakers rated the stimuli as highly natural and accurately (>95%) identified the tone categories.

Procedure

In all experiments, each auditory stimulus was presented with the following written prompt: “Which category? (Press the number key).” Participants generated a response by pressing one of four buttons on a keyboard labeled “1,” “2,” “3,” or “4,” corresponding to T1, T2, T3, and T4, respectively. Feedback was displayed for 1,000 ms after the response, depending on the experimental condition: immediately or delayed by 500 or 1,000 ms. The content of feedback varied depending on the accuracy of the response (“Correct/No”) and the experimental condition, in which full feedback informed the participant of the correct answer regardless of the accuracy. The stimulus–response–feedback sequence made up a single trial. One block consisted of a randomized presentation of all 80 stimuli. Six blocks were presented, yielding a total of 480 trials per participant. In Experiment 1, the response-to-feedback interval was immediate or delayed (1 s), always with full information. In Experiment 2, feedback information content was minimal or full, with a fixed response-to-feedback interval of 500 ms. In Experiment 3, the order of presentation of talker information was manipulated. In the mixed talker condition, the stimuli were presented in a random sequence. In the blocked talker condition, the stimuli were randomized only in terms of syllables and tones but ordered in terms of talker information. The feedback in Experiment 3 was always immediate and minimal (Fig. 2). In short, for all experiments, six blocks of 80 trials each were presented to the participants. The order of talkers was completely randomized for Experiments 1 and 2. In Experiment 3, in the blocked condition, the sequence of talkers (n = 4) was randomized. Within a talker sequence, participants listened to all stimuli (also randomized) produced by the talker (n = 20).

Results

Category learning over training trials

Figure 3 shows the proportion of participants who made a correct category response (sliding window of 80 trials) across conditions. Visually, reflexive and reflective conditions are more equivalent in the beginning than toward the end of the experiment, where the reflexive conditions consistently lead to improved learning progress. This inspection was corroborated with a statistical analysis detailed in the next section.

Fig. 3
figure 3

Category learning curves across reflexive versus reflective conditions in all three experiments: a Experiment 1, feedback delay (immediate vs. delayed); b Experiment 2, feedback information (minimal vs. full); c Experiment 3, talker variability (mixed vs. blocked). Plotted in solid bold lines are the proportions of correct responses across participants within each condition over the course of learning. The black lines denote the reflexive conditions, and the red the reflective conditions. For purposes of visualization of trial-by-trial data, each point in the line denotes the average number of correct responses in a sliding 80-trial window. For trials preceding the 80th trial, cumulative averages were used. Plotted in thin lines are the ranges of standard error of the averages used in the sliding windows. Visual assessment of the learning curves suggest that both conditions result in equivalent degrees of category learning toward the earlier phase of experiment but that the reflexive condition leads to greater learning than does the reflective condition toward the later phase of the experiment. This pattern is consistent across all three experiments

Full size image

Data analysis

For each participant, response to each trial was coded as “correct” or “incorrect.” A mixed logit analysis was conducted to predict the log odds of producing a correct response, using the lmer program with binomial logit link (Bates, Maechler, & Bolker, 2012). The dependent variable was set as the “correct” or “incorrect” outcome of each response. The fixed effects of interest were the between-subjects condition (reflexive vs. reflective), trial number (increasing from 1 to 480; mean-centered to 0 and divided by 100), and their interaction term. The model was corrected for by-subject and by-item random intercepts, which was the most complex model as justified by the data (p < .05):

$$ \mathrm{Outcome}\sim \mathrm{Condition}\times \mathrm{Trial}+\left(1\Big|\mathrm{Subject}\right)+\left(1\Big|\mathrm{Item}\right). $$

Experiment 1: Immediate versus delayed feedback

The trial effect was significant (p < .0001; each successive trial increases the probability of an accurate response). The trial × condition interaction was significant (p < .0001; each successive trial increases probability of an accurate response, more for the reflexive-immediate than for the reflective-delayed condition). The condition effect was not significant (see Table 1).

Table 1 Effects of feedback delay on category learning
Full size table

Experiment 2: Full versus minimal feedback

The trial effect was significant (p < .0001; each successive trial increases the probability of an accurate response). The trial × condition interaction was significant (p < .0001; each successive trial increases probability of an accurate response, more for the reflexive-minimal than for the reflective-full condition). The condition effect was not significant (see Table 2).

Table 2 Effects of feedback information on category learning
Full size table

Experiment 3: Mixed versus blocked talker presentation

The trial effect was significant (p < .0001; each successive trial increases the probability of an accurate response). The trial × condition interaction was significant (p = .021; each successive trial increases the probability of an accurate response, more for reflexive-mixed than for reflective-blocked condition). The condition effect was not significant (see Table 3).

Table 3 Effects of talker variability on category learning
Full size table

Discussion

We hypothesized that the reflexive learning system is optimal for adult speech category learning. The results from Experiments 1–3 strongly support our hypothesis. All three experiments show that training manipulations that targeted the reflexive learning system enhanced learning, relative to those that targeted the reflective learning system.

In Experiments 1 and 2, we examined trial-by-trial feedback training manipulations. Both experiments yielded a significant interaction between trial and training manipulation, such that accuracy increase over trials was greater for manipulations that targeted the reflexive learning system. In Experiment 1, we found that immediate feedback enhances tone learning toward the end of training, relative to delayed feedback. As per the dual-system models, immediate feedback is a critical requirement of the dopamine-mediated reflexive learning system, but not reflective learning. Therefore, delaying feedback even by just 1 s disrupts the dopamine-mediated training signal for the reflexive learning system, allowing control to pass to the reflective learning system. In Experiment 2, we found that minimal feedback enhances tone learning toward the end of training, relative to full feedback. Since the full feedback provides information not only about the correctness of the response, but also about the correct category membership, interpretation of this result is counterintuitive without the consideration of the dual-system perspective; full feedback promotes hypothesis generation and testing, which prevents transfer of control from the reflective system to the reflexive system. Therefore, Experiment 2 also supports the hypothesis that the reflexive system is optimal for speech category learning.

In Experiment 3, we manipulated the order of talker presentation: randomly mixed or systematically blocked. This experiment yielded an interaction between trial and training manipulation, such that accuracy increase over trials was greater for the mixed talker condition, relative to the blocked talker condition. This is consistent with our prediction that the mixed condition targets the reflexive learning system by preventing talker-dependent hypothesis generation and testing by the reflective system. As a result, participants in the mixed talker condition are led to rely more on relatively talker-invariant acoustic cues (e.g., pitch direction) than on talker-variant cues (e.g., pitch height). These results are consistent with those of the L1 speech acquisition literature that show that high variability can guide category learning by allowing associations between more invariant cues and the category structure (Apfelbaum & McMurray, 2011; Rost & McMurray, 2009), but not consistent with the results of a previous study that found greater Mandarin tone learning in the blocked talker condition than in the mixed talker condition (Perrachione et al., 2011). These contradictory findings, we believe, may be due to methodological differences. The high degree of variability in natural Mandarin tone categories was somewhat reduced in the previous study, where only three of the categories (T1, T2, and T4) were presented and the tone contours interpolated linearly. In contrast, our study utilized the full category structure with the natural tone contours retained (Fig. 1). We have argued earlier that multidimensionality is what makes speech categories hard to verbalize and, therefore, optimally learned through the reflexive system. In contrast, the reflective system is more likely to suffice as the complexity of the category structure is reduced.

In adult speech learning studies, significant category learning is evidenced without feedback, consistent with unsupervised learning models. The present study did not include a no-feedback condition that could help tease apart the relative contribution of unsupervised implicit learning in mediating category learning success. However, category learning is substantially more efficient with feedback (Goudbeek et al., 2008; McClelland et al., 2002). Our theoretical approach does not rule out the statistical learning processes operative within the primary and secondary auditory cortices. Rather, our results demonstrate the operational specifics of the domain-general feedback-based learning system that is optimal for learning speech categories. Although neuroimaging studies have implicated the neostriatum in speech category learning tasks, the role of the reflexive system has not been systematically examined. This is despite that fact that animal models clearly demonstrate direct, extensive, and many-to-one connectivity between the primary/secondary auditory cortex and the reflexive systems, suggesting a distinct neurobiological plausibility for a substantial role for this circuitry (Petrides & Pandya, 1988; Yeterian & Pandya, 1998). Future research should systematically examine the relative contribution of various forms of perceptual and learning processes to speech category learning, as well as use other category structures.

In summary, our results demonstrate that speech category learning is optimally learned by the reflexive system. Our results offer practical implications for the development of optimized training approaches that can target the reflexive learning system. Specifically, we hypothesize that for speech categories, learning can be optimized by including minimal and immediate feedback and high trial-by-trial talker variability.