sounds or suprasegmentals.
Instructed L2 Pronunciation Development
While limits on ultimate attainment in naturalistic learning contexts are well-established, instructed L2 pronunciation can provide an opportunity to re-orient learners’ selective perception towards phonetic cues that they have learned to ignore. Despite this fact, production-oriented approaches to teaching pronunciation have long-dominated the field (e.g., Celce-Murcia et al., 2010; Lyster et al., 2013; Thomson & Derwing, 2015). In the 1970s, Audiolingualism used spoken models, but only to introduce production rehearsal activities (Brown & Lee, 2015). In the 1980s, Communicative Language Teaching (CLT) largely neglected pronunciation instruction, focusing almost exclusively on communicative processes to the exclusion of form. While CLT appealed to adult L2 learners’ desire to communicate as quickly as possible, it largely recreated the naturalistic conditions under which L2 pronunciation is most resistant to change. In more recent years, while pronunciation instruction has regained a position of importance, instruction continues to be predominantly production-focused (e.g., Saito & Lyster, 2012).
Since we know that speech perception plays such a foundational role in accurate speech production, why is this not reflected in popular L2 pronunciation teaching methods? One reason for this disconnect is that L2 speech perception research is largely inaccessible to language teachers (Thomson, 2018a). It is typically published in technical journals and largely reports on laboratory-based studies, which may appear to lack relevance to the real world. Unfortunately, by ignoring this important research, pronunciation specialists are unprepared to teach in a way that addresses the cognitive-developmental features of pronunciation associated with speech perception (Derwing & Munro, 2015).
In laboratory studies, changes in L2 speech perception happen consistently and rapidly (Sakai & Moorman, 2018). Furthermore, perceptual training typically leads to improvement in production, albeit more slowly than in perception. The perceptual training technique with the most empirical support is High Variability Pronunciation Training (HVPT) (see Thomson, 2018a). This technique is based on evidence that to be effective, perceptual training needs to include variation in terms of the number of talkers whose voices comprise training stimuli and in the number of phonetic contexts (or words) in which target sounds are presented. Training on a single talker does not generalize to perception of the target sounds spoken by new talkers. Presentation of target sounds in multiple phonetic contexts is important because learning to perceive a sound in one context (e.g., the vowels in “hit” and “heat”) rarely transfers to perception of the same sound in different contexts (e.g., to the vowels in “sit” and “seat”) (Thomson, 2016, 2018a). There is also evidence that training target sounds using nonsense syllables/words is initially more effective than training using real words (Thomson & Derwing, 2016). This may be due to the ability to orient learners’ attention to sounds in nonsense words, without competition from meaning (Guion & Pederson, 2007). HVPT typically presents training stimuli by computer or mobile application. Learners hear training tokens and must respond by clicking/tapping on a symbol, letter, or word representing the sound that they just perceived. They then receive feedback on the accuracy of their responses. There is limited direct evidence for how to use HVPT to train suprasegmental features, but there is some indication that it would have a similarly beneficial effect (Thomson, 2018a).
The cognitive mechanisms underlying the benefits of HVPT over low variability training are not fully understood. One possibility is that all cognitive categories, by their very nature, contain variation. This means that learning a new category necessitates learning about the distribution of sounds that can occur within that category. Another possibility is that the use of multiple talkers maximizes the potential for a given L2 learner to encounter at least some tokens of target sounds that do not automatically assimilate to pre-existing L1 categories. There is evidence that L2 learners are more apt to recognize English vowels as belonging to a new category if those vowels were produced by a talker whose productions are acoustically distant from any confusable L1 vowel categories (Thomson, 2007). It remains unclear whether there is an optimal amount of variability. Programs claiming to be HVPT have used between 2 and 30 talkers, for example, and vary widely in the number of phonetic contexts utilized (Thomson, 2018a). It seems unlikely that using two talkers provides optimal variability, but using increasingly larger numbers may result in diminishing returns or make learning more difficult (Thomson, 2018a).
In addition to variation in stimulus talkers and phonetic contexts, the nature of corrective feedback (CF) has also been shown to impact perceptual learning. While corrective feedback of some sort is a feature of all HVPT training, Lee and Lyster (2017) explored what role specific types of correct CF play in the transfer of perceptual training to production. Their HVPT experiment tested CF in four conditions. One group received feedback only that their response was incorrect, but was not told what the correct response should have been. Three other groups received feedback when they were wrong, but also received additional input, either by hearing an example of the target item again, hearing an example of the non-target item they had selected, or hearing both the target and non-target sounds. Only the groups that heard either the target form or the non-target form as part of CF displayed transfer to production. Those that did not hear either target or non-target items, or those that heard both, did not improve in production. In sum, Lee and Lyster’s study confirms that drawing learners’ attention to errors, either through negative or positive evidence, contributes to learning, just as it does in L1 pronunciation development.
Two other approaches to perceptual training are also worth noting, although neither has a sufficient evidence base to support its widespread adoption. Underlying both seems to be an assumption that accurate perception of one’s own productions (i.e., self-perception) has a facilitative effect in ultimately matching what one perceives with what one produces (Baker & Trofimovich, 2006; Borden et al., 1983). While neither is a purely perception-oriented technique, both include a perceptual component through the use of imitation. Rojczyk (2015) instructed Polish learners of English to imitate English-accented Polish sentences, targeting particular English consonants. This practice may have had the effect of orienting learners’ attention to phonetic information produced with their own voices in their L1. The researcher found that imitation of English-accented Polish led to positive transfer in their L2 English pronunciation. Other researchers are testing what they call a “Golden Speaker” approach to making perceptual learning easier (Ding et al., 2019). This is based on a belief that there are ideal voices from which particular learners can best develop L2 speech perception. The Golden Speaker web-based application maps a learner’s own voice quality onto the correct pronunciation of target sounds produced by a native speaker. The system then generates training stimuli that sound like the learner’s voice, but without segmental errors. It is assumed that this will make it easier for learners to attend to those parts of the acoustic signal that are distinct from their own voices, because it simulates a vocal tract size and shape that is exactly the same as the learners’ own, but with an articulation model that is native speaker-like. To the extent to which these alternative approaches to L2 perception training work, they may be preferable to HVPT, which is more labor-intensive since it requires the accumulation of training stimuli from multiple talkers.
Pedagogical Implications
While HVPT is effective in re-educating selective perception across the lifespan, and encourages progress beyond the WMO (Thomson, 2012a), it is not particularly easy to implement. Until user-friendly HVPT training platforms become more widely accessible, classroom pronunciation instruction will remain the most common mode of delivery. Even when HVPT or other digital applications are used for perception training, they should be seen as a complement to classroom-based instruction, not as a replacement. Ultimately, learners still need opportunities for live interaction to move beyond highly controlled L2 pronunciation skills and towards automatic L2 speech production. In designing classroom instructional tasks, teachers should take note of key findings in the L2 speech perception literature described above, first, by allowing more instructional time for perception-oriented training. Short batteries of perceptual instruction trigger improvement in both perception and production
