Individual variation in performing reading-aloud speech among deaf speakers

3.1 Linguistic variable

This study looks at /s/ and /ʂ/ in Taiwan Mandarin. /s/ and /ʂ/ are in an alveolar-retroflex phonemic contrast which is merging in the direction of the alveolar (Kubler 1985; Tso 2017). This article focuses on only fricatives but not affricates in an alveolar-retroflex contrast, for sibilant fricatives are auditorily much more challenging for deaf speakers to acquire, thereby receiving much attention from both deaf speakers and speech pathologists (Yang et al. 2017).

While the retroflex consonant is known as 捲舌音 juǎnshé yīn ‘tongue-curling sounds’ in Mandarin phonics in Taiwan, Taiwan Mandarin speakers who demonstrate a clear alveolar-retroflex distinction produce the retroflex consonant by elevating the tongue tip and blade (Chiu et al. 2020). The retroflex is produced with a constriction at the alveolar zone (Chen and Chang 2015). Acoustically, if the contrast is maintained, retroflex sibilants receive lower spectral centre of gravity (CoG) than alveolar sibilants do. A fronted sibilant leads to a higher CoG.

Full retroflexion in /ʂ/ is ideologically loaded. Mandarin speakers from China are perceived by Taiwanese people to demonstrate full retroflexion. For instance, Taiwanese pop singers do full retroflexion to style Chineseness (Lin and Chan 2022). As the imagined standard accent of Mandarin is believed to be oriented to a Beijing accent, there is a decline of 200–300 Hz in the CoG of /ʂ/ when hearing speakers shift to read speech from interview speech (Tso 2017). In contrast, stylistic variation usually does not involve the alveolar sibilant: if anything, speakers might hypercorrect it as a retroflex sibilant (Kuo 2018).

Variation in the sibilants is auditorily inaccessible for many deaf speakers. As the alveolar-retroflex contrast is highly socially meaningful, it is one of the targets in speech therapy when therapists work with deaf children. That is, it is expected to see deaf persons who engaged in medicalization of their speech more aligned with hearing people’s community pattern of realizing the contrast. For instance, the 8-year-old boy Kuo knows what specific phonetic variants are considered prestigious by hearing people in the context of reading aloud, after training. In contrast, others who engaged less in speech medicalization may have the idea that they are expected to shift to a speech style which can embody able-bodiedness when reading aloud, but they might not be familiar with how hearing people perform prestige.

3.2 Participant and data collection

A total of 10 deaf participants are included in the present analysis (Table 1), all recruited from Taipei through advertisements in public Facebook groups between December 2018 and January 2019. The average age is 24.3 (max = 44; min = 18). In the recruitment text, it’s specified that this study aims to record and describe deaf speech without evaluating the correctness of deaf speech.

Only one participant (Huei) identifies as culturally deaf because she is actively learning Taiwan Sign Language and connecting herself to deaf signing communities.

The researcher is a hearing man from Taipei, aged 25 at the time of interview. All the data used were recorded by a Zoom H5 (sampling rate = 44.1 kHz), with built-in microphone. All names are pseudonyms offered by the participants.

Each participant participates in a reading task and an interview. The interview includes questions including various aspects of how they deal with deafness in everyday life.

To investigate stylistic variation, read speech and interview speech are compared. For read speech, participants are invited to read two passages out loud. One of the passages is about a royal story in Qing Dynasty of China, and the other one is about an allegory of the relationship between deaf people and hearing people. The stylistic variation between the two passages is examined in another article of this project (Wan in press).

There are 47 tokens with /ʂ/ and 10 tokens with /s/ in read speech. For interview speech, I used the /ʂ/ and /s/ tokens produced when participants responded to my first interview questions, including “which identity label do you identify as?” “how do you see your relationship with Taiwan Sign Language?” and “could you tell me about your medical history of deafness?” The first two questions are answered before the reading task, and the third question is answered after the reading task.

CoG has been observed to be a robust spectral moment that indicates the articulatory setting of realizing retroflex sibilants (Chang and Shih 2015; Chiu et al. 2020). Thus, this study uses CoG as an indicator of the tongue movement in producing sibilants. The midpoint CoG value of the frication interval of each token is automatically extracted by a Praat script (Reetz 2020), using a 25 ms Gaussian window (Boersma and Weenink 2020). A high-pass filter is set at 1,000 Hz to eliminate low frequency noise and co-articulatory voicing from the surrounding segments (Chang and Shih 2015). Tokens whose duration is less than 30 ms are excluded.

While the use of retroflexion among hearing speakers may be affected by the ideological differentiation between Chineseness and Taiwaneseness (Brubaker 2012; Chen 2018), data about political identity were not collected from the participants. The interview is completely about deafness and disability. Individual political identity is not primed at all, and the current political climate does not emerge as a theme in the interview either. Although one of the passages is a Chinese royal story from 300 years ago, it is not about contemporary China. In addition, the participants explicitly point out that they are not familiar with the story and are not fans of Chinese harem dramas. In this article, I therefore focus the analysis of stylistic variation on disability-relevant biographical factors.

4.1 Statistical analysis

Linear mixed effects models were fit into the CoG data by the lme4 package (Bates et al. 2015) in R (R Core Team 2019). Data points whose CoG is not within two standard deviations from the mean are excluded respectively from the alveolar fricative dataset and retroflex fricative dataset. As there are only four tokens followed by [e] (all from interview speech), they are excluded from the data. The two datasets are then merged, resulting in a total of 2,238 observations (Tables 2 and 3).

Models were expanded incrementally from the null model. Fixed factors were added one by one. Fixed effects of linguistic constraints (Table 4) considered during model selection included place of articulation (alveolar vs. retroflex), the following vowel, and log-transformed frication duration (Chang and Shih 2015). The following vowel has been observed affecting the CoG of frication (Chang and Shih 2015; Chiu et al. 2020). The effect of prosodic prominence was already considered in the factor of frication duration (see Chang and Shih 2015), so no separate factor of prosodic prominence was considered.

Gender balance was not achieved during data collection. Among the 10 participants, there are eight women and two men. Gender and its interaction with register or place of articulation failed to achieve statistical significance, nor did they improve the model fit. This was also the case for age (continuous). Therefore, gender and age are not further discussed in the following statistical analysis. Educational levels of the ten participants are relatively homogenous. There is no theoretical motivation to believe educational level as a binary variable (bachelor’s vs. master’s) has any effect on the use of the linguistic variables.

As mentioned, whether a participant underwent speech therapy or private speech training may have an influence on how they engage with the target variables. Thus, it is considered in statistical modeling here, referred to as the factor of “medicalization” (Yes vs. No).

Interaction terms among register, place of articulation, and medicalization were subsequently added to the model where possible. Likelihood ratio tests were run to determine whether the model fit of an expanded model is improved. By-speaker intercept was included as a random effect. By-speaker random slopes were included where possible. Model expansion continued until incorporating another independent variable did not improve the model fit.

4.2 Findings

The best-fitting model is CoG ∼ Vowel + Sibilant duration + Place of articulation ⋆ Medicalization ⋆ Register + (1 + Place + Register|Speaker) (Table 5).

As predicted, sibilant duration is positively correlated to CoG (Chang and Shih 2015). Tokens followed by [o] or [u] have significantly lower CoG than those followed by [a], consistent with previous research (Chang and Shih 2015; Chiu et al. 2020).

Place of articulation is not a significant effect for non-medicalized speakers. In contrast, the significant interaction between medicalization and place of articulation indicates that, at the reference level (i.e., read speech), the retroflex fricative receives lower CoG than the alveolar fricative does, among medicalized speakers; this effect is significantly reduced in the interview speech (as indicated in the significant three-way interaction).

As shown in Figure 2, among speakers who underwent speech therapy or training, the retroflex sibilant becomes more retroflexed in read speech, compared to the spontaneous speech. In contrast, the alveolar fricative barely shifts, which is also indicated by the lack of significant interaction between register and medicalization. This is consistent with how hearing people in Taiwan shift the CoG of the retroflex sibilant rather than the alveolar sibilant when reading aloud (Tso 2017).

Figure 2:

The interaction among register, medicalization (yes/no), and place of articulation.

Among the non-medicalized speakers, there is no significant change in the CoG across register, but we see a tendency for raising the CoG in read speech, regardless of the place of articulation. As there is no group-level effect of register in the non-medicalized speakers, it is useful if we know whether particular speakers among them raise the CoG in read speech to a larger extent. The best-fitting model has already considered the factor of medicalization. We need a model which only controls for the linguistic constraints, showing how the speakers exactly engage with the reading-aloud register.

I fitted a model without the factor of medicalization to the data: CoG ∼ Vowel + Sibilant duration + Place of articulation * Register + (1 + Place + Register|Speaker). The effects of following vowel, and sibilant duration remain in the same direction (Table 6). The model summary shows that there is no significant fixed effect of register. The two-way interaction is not significant, either.

After we only control for the linguistic constraints, random slopes show what individual speakers do in read speech. A growing amount of sociolinguistic research has made use of random effects to explore how the same linguistic effect (including register) varies across individuals (e.g., Drager and Hay 2012; Hall-Lew 2013; Forrest 2015; Schleef 2021).

In Figure 3, we can see the individual-level deviations from the group-level fixed effects of register (which does not achieve statistical significance) and place of articulation (which achieves statistical significance) (Drager and Hay 2012). Four speakers – Sandy, QPM, McCrispy, and Huei – increase alveolar constriction in read speech. Their personal slopes of register have 95% confidence intervals which do not overlap 0, which means they increase alveolar constriction to do the reading-aloud accent.

Figure 3:

Estimated deviations from the group-level fixed effects of place of articulation and register (non-medicalized participants are underlined).

In addition, not every speaker demonstrates the same CoG difference between places of articulation. Sandy, McCrispy, and Kai demonstrate personal slopes of place of articulation with 95% confidence intervals which do not overlap 0. The difference between retroflex and alveolar sibilants is less distinct than the one indicated by the group-level effect. With the coef function in the stats R package, the predicted effect of place of articulation for individual speakers is extracted (considering both the fixed slope and random slope) (Figure 4).

Figure 4:

Predicted effects of place of articulation for individuals.

4.3 Analysis

We have identified the variation pattern. What motivates the style-shifting we have seen so far? It is not surprising to see speakers who engaged in speech medicalization shift to hearing people’s community pattern in read speech. The only exception is Huei.

Huei is the only person who identifies as culturally deaf. She received a cochlear implant and engaged in speech therapy as a child. She also follows the hearing people’s community pattern to articulate the phonemic contrast – retroflex sibilants receive lower CoG than alveolar sibilants do. However, she diverges from the hearing people’s community pattern of engaging with the reading task. That is, in the self-conscious reading task performed in front of a hearing researcher, Huei uses her first-learned style rather than a corrected speech style added to her repertoire later in life. While it seems that Huei might exploit her first-learned style (the ‘deaf accent’) to perform a certain culturally deaf identity, such an account presumes that individuals utilize linguistic resources to index a group identity (Sharma 2021). In the interview, Huei points out that she does not favor the term ‘deaf accent’ because she thinks every person talks differently but “naturally” and that the ways deaf people speak shouldn’t be highlighted, lumped together, or studied by speech pathologists. She also emphasizes that she does not speak Mandarin in a way that follows what’s described by speech pathologists as deaf speech and points out that many hearing persons have said her pronunciation is standard. In other words, to some extent, Huei rejects any attempt to list her speech style as an example of ‘deaf speech’ and takes the “real me” stance that performs how her ‘natural’ way of speaking is a legitimate way of speaking. Thus, in the reading task, Huei does not perform a deaf identity so much as present a personal ‘natural’ way of speaking as a kind of embodied able-bodiedness, which is disassociated from hearing people’s way of doing able-bodiedness (i.e., the idea of a standard speech). Note that we should not too optimistically take Huei’s stylistic practice here to mean that deaf speakers have much social agency in negotiating with ableism. Compared to reading-aloud in a context like the National Language Competition, a reading task in the sociolinguistic interview is not co-present with a large audience and a dominant linguistic market (Bourdieu 1977) where a particular linguistic variety is explicitly valued the most. Thereby, this possibly creates an atmosphere where marginalized speakers like deaf people can exhibit more agency in negotiating with the hegemonic standard language ideology which is entangled with able-bodiedness.

Sandy, QPM, McCrispy and Kai did not receive speech therapy or training. Kai is the only one of them who utilizes retroflexion in read speech. While Kai did not receive speech therapy, he has a strong metalinguistic awareness of the alveolar-retroflex contrast. According to Kai, he could not hear the difference between the retroflex and the alveolar sibilants, so he usually does not differentiate the two sibilants in speech production (as shown in Figure 4). He also explicitly points out that the alveolar variant is his unmarked variant because he knows hearing people use the alveolar variant in conversational speech. For Kai, the retroflex sibilant as a marked variant is his later-learned speech style. His metalinguistic awareness may mobilize retroflexion in read speech.

In contrast, Sandy, QPM and McCrispy utilize increased alveolar constriction in read speech. Even if they acquired the phonological representation of the alveolar-retroflex contrast (Figure 4), they utilize a fronted sibilant to a larger degree when shifting to read speech.

First, we need to make sure the three speakers make use of spoken Mandarin “as a measure of proximity to hearingness and thus ability” (Henner and Robinson 2021b: 18). Gafter (2021) points out that not every person has the same stylistic goal in a reading task. One possibility is that the three speakers might distance themselves from hearingness in the reading task. However, in the interviews, all the three participants expressed their positive stance towards hearing speech and found ‘deaf accents’ not desirable. For instance, in the interview, McCrispy highlights a 濁 zhuó ‘turbid’ quality in her own voice (Extract 1; my own translation):

Extract 1 (M: McCrispy; R: Researcher)

In Extract 1, McCrispy uses the term zhuó to locate a quality of her speech that she wants to change (line 7). The “turbid” quality is undesirable, and it is contrasted with a “clean” quality (line 20). McCrispy reports that she acquired deafness at the age of 3. She believes her first-learned speech style is relatively clean, largely based on her perception of how hearing people talk. This “clean” quality may be picked out from the ideological position of able-bodiedness (cf. Kusters 2020). In a further explanation, McCrispy points out that the “turbid” quality leads to the perception that “all the words stick together”, resulting in an unclear speech style. Without experiences with speech medicalization, a fronted sibilant might be a “semiotic hitchhiker” (Mendoza-Denton 2011), employed by McCrispy as a part of a “clean” (able-bodied) speech. McCrispy might shift to her first-learned speech style in the reading task to perform her once-owned hearingness.

Yet, what leads to the link between an increased alveolar constriction and able-bodiedness? McCrispy does not explicitly comment on it, whereas Sandy does. Sandy is the person who raises the CoG of the sibilants the most among the three speakers. She mentioned she was told by a teacher at her elementary school that she needed to “expel a breath” (吹氣 chuīqì) when pronouncing the sibilants. This refers to the hiss in /ʂ/. Before Sandy learned this speech style, she realized it as an affricate [tʂ], for the hiss is not auditorily accessible to Sandy.

She takes the word 老師 lǎoshī ‘teachers’ for example, where the second syllable starts with /ʂ/. She repeats the phrase twice: “老師的師” (‘lǎoshī’s shī’) (Figure 5). When she singles out the second syllable, the sibilant receives a higher CoG value than when it is unfocused; sibilant duration does not change. That is, to emphasize the hiss, she advances her tongue. The fronted sibilant is actually treated as the standard variant in her stylistic repertoire. In the reading task, Sandy shifts to her later-learned speech style to perform able-bodiedness.

Figure 5:

CoG and duration of the sibilant /ʂ/ in the word lǎoshī by Sandy.

As a side note, the observation can advance our understanding of how a fronted sibilant is usually treated as the embodiment of a harsh speech style (Calder 2019; Eckert 2017). Eckert (2017: 1,201) points out that “language is embodied, and that sound symbolism builds on that embodiment”. She observes that “attenuating fortition attenuates the meaning” of a hiss: when the tongue is retracted, a hiss becomes less harsh and sharp. Previous research explores the social qualities associated with sibilant variants mainly from the perspective of auditory perception. For example, a higher CoG is iconic of smallness, and further ideological associations can be made with the quality of smallness (Eckert 2017); or, such an “extreme acoustic manifestation” generates the feelings of sharpness, harshness, and thereby fierceness (Calder 2019). Here, I argue from a perspective of articulatory settings (e.g., Pratt 2019; Podesva 2021). An advanced tongue leads to a narrower constriction area in the alveolar zone, which focuses the stream of air more intensely, resulting in greater air particle velocity (Stevens 1998). Sandy is not aware of how hearing people move the tongue to realize a standard retroflex sibilant. Thus, the embodied feeling of the link between a fronted sibilant and a stronger hiss is not overridden by any metalinguistic discourse informed by medical professionals. In fact, a sibilant with a relatively low CoG is more perceptible for deaf people (Maniwa et al. 2008). If Sandy attempted to produce a sibilant noise that sounds clearer to her, the tongue should have been retracted. That is, it is the feeling induced by the articulatory gesture, rather than an acoustic quality, that is prioritized here when the ‘standard’ speech is performed.