Perception of incompletely neutralized voicing cues in word-final obstruents: The role of differences in production context

1 Introduction

Traditional accounts of word-final obstruent devoicing assume that the phonological voicing contrast is not preserved at the phonetic level, with word pairs such as Rad ‘wheel’ and Rat ‘advice’ in German and код /kod/ ‘code’ vs. кот /kot/ ‘cat’ in Russian being indistinguishable in both production and perception (Kiparsky 1976, among others). At the same time, experimental studies have repeatedly shown that speakers of neutralizing languages maintain acoustic differences between phonologically voiced versus voiceless final obstruents and that listeners can often identify the voicing setting of such segments at an above-chance level (e.g., Port and O’Dell 1985; Warner et al. 2004). The current study investigates how such identification responses vary depending on whether the perceptual stimuli were recorded using reading vs. non-reading procedures and with vs. without full minimal pairs in the experimental list. The goal is to determine how differences in production context affect the extent to which the phonological contrast is neutralized in perception.

Production data from languages with final devoicing, including German, Dutch, Afrikaans, Polish, and Russian, often show reliable differences between the surface forms of phonologically voiced versus voiceless obstruents (e.g., Chen 1970; Charles-Luce 1985; Port and O’Dell 1985; Slowiaczek and Dinnsen 1985; Tieszen 1997; van Rooy et al. 2003; Piroth and Janker 2004; Smith et al. 2009; Dmitrieva et al. 2010; Rӧttger et al. 2011; Kharlamov 2014; Röttger et al. 2014). This phenomenon is traditionally referred to as ‘incomplete neutralization’, and it tends to affect the acoustic parameters of consonantal duration, glottal pulsing, and preceding vowel duration. Phonologically voiced consonants usually show significantly shorter durations, more glottal pulsing, and longer preceding vowels compared to their voiceless counterparts (e.g., all 3 parameters demonstrate significant differences in Port and O’Dell 1985; van Rooy et al. 2003).

Incomplete neutralization has also been attested in the domain of perception, where the rates of voiced responses vary depending on the phonological voicing setting of the final consonant despite its phonetic devoicing (Port and O’Dell 1985; Port and Crawford 1989; Slowiaczek and Szymanska 1989; Warner et al. 2004; Röttger et al. 2011). For example, during forced-choice two-alternative identification tasks, speakers of German have shown above-chance-level performance on voicing, with the majority of phonologically voiced but phonetically devoiced segments being attributed to the voiced category (e.g., Port and O’Dell 1985; Port and Crawford 1989; Röttger et al. 2011). Comparable results have also been reported for other devoicing languages, including Dutch (e.g., Warner et al. 2004) and Polish (e.g., Slowiaczek and Szymanska 1989), as well as for other types of perceptual tasks, such as discrimination experiments (e.g., Matsui 2011) and rating tasks (e.g., Ernestus et al. 2007a). This suggests that listeners perceive the partial cues to voicing and interpret them as evidence of the target segment being [+voiced].

When voicing effects are observed in production data, they are usually attributed to speakers’ producing incompletely neutralized consonants because they are aware of these segments being voiced at the phonological level or because they know that the same consonant is articulated as voiced when non-final in morphologically-related forms (among others, Port 1996; Port and Leary 2005; Ernestus et al. 2007a, Ernestus et al. 2007b). However, incompletely neutralized differences are also known to be motivated in production by speakers’ exposure to orthographic representations during word-reading tasks and inclusion of minimal pairs among the stimuli (among others, Fourakis and Iverson 1984; Mascaró 1987; Jassem and Richter 1989; Manaster Ramer 1996a, Manaster Ramer 1996b; Warner et al. 2004; Warner et al. 2006; Iverson and Salmons 2011; Kharlamov 2014). For example, Fourakis and Iverson (1984) found incompletely neutralized differences in German in the tokens elicited during a word-reading task but not in the items recorded during an oral task. Kharlamov (2014) reported that robust differences in glottal pulsing between phonologically voiced versus voiceless stops and fricatives in Russian were found during word-reading but not picture-naming/word-guessing, and when the list of test items contained minimally-contrasting word pairs (e.g., /kot/ ‘cat’ ~ /kod/ ‘code’) but not when such pairs were excluded. This shows that both orthography and lexical competition can affect the production of voicing in obstruents, likely by increasing speakers’ awareness of the situational importance of voicing and encouraging the production of prominent acoustic differences between voiced and voiceless obstruents in the parameter or parameters relevant for a given language (e.g., glottal pulsing for Russian speakers, segmental duration for German speakers).

Considering that the previous perception findings have been based on (i) the stimuli recorded in word-reading tasks with voicing-based minimal pairs included in the experimental list (e.g., Port and O’Dell 1985; Port and Crawford 1989; Slowiaczek and Szymanska 1989; Warner et al. 2004), (ii) word-reading tasks without minimal pairs among the stimuli (e.g., Ernestus et al. 2007a), or (iii) oral tasks with minimal pairs among the stimuli (e.g., Jassem and Richter 1989; Matsui 2011; Röttger et al. 2011), listeners’ apparent ability to correctly identify intended voicing may then be due to their reliance on those hyper-articulated differences that are present in the stimuli as a result of speakers’ exposure to either written forms or lexical competition or both. Moreover, many previous perception studies limited their scope to monosyllabic words ending in coronal stops (e.g., Port and Crawford 1989; van Rooy et al. 2003; Warner et al. 2004) and relied on small samples of speakers or stimulus items, such as testing 10 or fewer participants and/or stimulus items (e.g., Port and Crawford 1989; Kopkalli 1993; Matsui 2011; Röttger et al. 2011). However, the patterns that exist at the level of individual speakers or specific words may not necessarily be representative of group-level behaviors and the entire lexicon, and factors such as word length and consonantal place of articulation are also well-known to play a role in the production of the voicing contrast in obstruents (among others, Gamkrelidze 1975; Ohala 1983; Warner and Tucker 2011). As such, the previously reported cases of the preservation of the voicing contrast in perception do not necessarily indicate that perceptual neutralization is incomplete in the language in general. Instead, it may be incomplete only for a subset of lexical items, specific obstruent types, and/or only the stimuli from those production contexts that involve speakers’ exposure to orthographic forms and lexical competition.

The study presented below investigates the exact role of differences in production context during the identification of voicing in Russian, a language that has so far received only limited attention in the experimental literature on the perception of incompletely neutralized contrasts. The study examines participants’ performance on the stimulus items recorded during word-reading vs. picture-naming/word-guessing tasks and in the presence vs. absence of minimal pairs in the experimental list. To ensure that the results are representative of (the local variety of) the Russian language in general, the findings are based on a large set of obstruents and large samples of participants and stimulus items.

2 Method

3 Results

4 Discussion

The current study examined how identification of underlying voicing in incompletely neutralized obstruents varies across stimulus items recorded during reading vs. non-reading procedures when minimal pairs are either present or absent among the stimuli. Results from Russian revealed that, overall, underlying voicing was perceptible with limited accuracy only. On average, between 42–44% of voiced obstruents and 26–30% of voiceless consonants were attributed to the voiced category. This shows a general bias for voiceless responses, which parallels the findings of the Polish study in Jassem and Richter (1989) and the Afrikaans investigation in van Rooy et al. (2003) that also found high proportions of voiceless responses to all final obstruents regardless of their phonological specification.

At the same time, identification responses also differed depending on which production context the stimuli came from. When perceptual tokens were originally recorded during word-reading and with minimal pairs included in the experimental list, responses to the two voicing categories differed by 21–25%. When the same items were created during a non-reading procedure and in the absence of minimal pairs among the stimuli, response rates showed differences of 6–10% only. Thus, identification responses were less consistent with phonological specifications for the tokens elicited in non-reading tasks and without minimal pairs among the stimuli. This shows an effect of differences in production context and suggests that identification responses were affected by the acoustic cue of glottal pulsing, which is the parameter that showed robust production differences during word-reading and in the presence of minimal pairs but either limited or no differences during non-reading tasks and in the absence of minimally-contrasting word pairs. The primary role of glottal pulsing is also corroborated by the findings of regression modeling, with vocal fold vibration being the most prominent predictor for plosive-final monosyllables and the only significant predictor for plosive-final disyllables and both monosyllabic and disyllabic fricative-final tokens.

However, even though the tokens recorded during non-reading tasks and without minimal pairs among the stimuli showed a strong voiceless bias, rates of voiced responses were still 6–10% higher for voiced than voiceless obstruents. This demonstrates that listeners are able to use not only the cues motivated by orthography or lexical competition but also the differences present in production even without speakers’ exposure to written forms or minimal pairs. This is further confirmed by the result of regression analyses in which closure/frication duration, release duration, and preceding vowel duration (which are the parameters that did not show any enhancement of production differences as a result of speakers’ exposure to orthography or lexical competition) were identified as limited but nevertheless significant predictors that accounted for 1.7–7.6% of variance. Note, however, that since only the steady-state part of the vowel was included in the rhyme-only perceptual stimuli, the exact nature of the interaction between voicing and preceding vowel duration remains to be investigated.

Taken together, these findings suggest that the apparently successful identification of underlying voicing in incompletely neutralized final obstruents can be attributed only in part to listeners’ taking advantage of the differences produced when speakers are exposed to orthographic representations and minimal pairs. When such cues are unavailable, listeners show a stronger bias for voiceless responses, yet the rates of responses to voiced versus voiceless obstruents are never identical. Thus, perceptual identification of voicing appears to involve both the prominent incompletely neutralized differences that are the result of speakers’ exposure to written forms and minimal pairs as well as the subtle incompletely neutralized differences that are not directly motivated by orthography or lexical competition.

For plosive-final tokens, perception results also revealed an effect of consonantal place of articulation, with velar stops showing greater consistency with underlying voicing. This parallels the findings of Port and O’Dell (1985) who also observed more accurate identification in the case of velar obstruents. As noted in the description of the stimuli, place of articulation did not have a significant effect in production on the degree of preservation of voicing. This means that listeners appear to be more sensitive to the presence of the same amount of glottal pulsing in velars than labials or coronals. Although the exact causes for this effect would need to be investigated independently in future research, one likely explanation comes from the aerodynamics of speech. Physiological reasons do not favor the production of glottal pulsing in velar stops. Velars are articulated with a smaller space between the glottis and the oral constriction, which results in a smaller volume of air being able to pass through the vocal folds during the closure stage of the obstruent (Gamkrelidze 1975; Ohala 1983; Ohala 1997). As voicing is less expected for velars in production, it may be more perceptually salient in velar plosives, and this may explain why Russian listeners performed better on velars than non-velars.

Finally, the findings presented above also have implications for the phonological accounts of final devoicing and incomplete neutralization. The data show that neutralization is not fully complete in perception for the stimuli recorded without speakers’ exposure to orthographic forms and minimal pairs, yet the observed differences only result in a slight weakening of the voiceless bias rather than a categorical shift in voicing judgments. As such, it seems unlikely that speakers produce incompletely neutralized differences because they intend to provide the listeners with cues to phonological voicing. If speakers’ goal were to disambiguate the potentially homophonous forms, voicing cues would be expected to have a major effect on the outcome of the categorization process across all production contexts and not only for the stimuli elicited during word-reading and with minimal pairs included in the experimental list. This suggests that such differences are not the result of a specific articulatory goal but more likely a by-product of various automated processes taking place during lexical access and speech planning, such as automatic co-activation of morphologically related forms (Ernestus and Baayen 2006; Goldrick and Blumstein 2006; Winter and Röttger 2011; Röttger et al. 2014). This seems especially likely considering the generally small magnitude of production differences in incomplete neutralization studies, such as the vowel duration differences of 3.5 ms in Warner et al. (2004) or the consonantal duration differences of 5 ms in Kharlamov (2014). During perception of non-laboratory speech, listeners also normally have access to non-acoustic (e.g., syntactic, semantic) cues to the voicing setting of the consonant, which further diminishes the importance of partially neutralized differences.

5 Conclusions

The study presented above examined the perception of incompletely neutralized voicing cues in word-final stops and fricatives in Russian. The goal of the investigation was to determine how identification of underlying voicing varied across stimulus items recorded using reading vs. non-reading procedures and with vs. without speakers’ exposure to minimal pairs. Results of a series of identification tasks revealed that listeners’ responses were more in line with underlying voicing for the stimuli elicited during word-reading and with minimal pairs included among the stimuli. However, significant differences in response rates were also found for the stimulus items from other production contexts. This shows that voicing judgments are influenced by the production context, yet perceptual neutralization can be incomplete even when the stimuli are recorded without speakers’ exposure to orthography and lexical competition.