Do Voice Acoustic Parameters Differ Between Bilingual English-Spanish Speakers and Monolingual English Speakers during English productions?

Abstract

Background:

In addition to language differences in fundamental frequency between bilinguals and monolinguals, studies have also included other acoustic parameters to analyze differences in voice production associated with the language spoken.

Aim:

To identify differences in voice acoustic parameters during English productions between monolingual and bilingual English speakers.

Method:

Exploratory cross-sectional study with two groups of subjects: monolingual English speakers (n= 40), and bilingual English-Spanish speakers (n= 13). Participants filled out a questionnaire and recorded one reading in English (second sentence of Rainbow passage “The rainbow is a division of white light into many beautiful colors”) under a “virtual-simulated” acoustic condition of No Noise and Medium Reverberation Time (0.8 sec).

Result:

Analysis by gender shows that monolingual speakers had higher fundamental frequency mode, and lower standard deviation of fundamental frequency compared to bilingual English-Spanish speakers. Bilingual male speakers had higher jitter and HNR than monolingual speakers. On the contrary, female bilingual speakers had lower jitter and shimmer than monolingual speakers.

Conclusions:

Speaking a second language may influence voice acoustic parameters, and therefore, should be considered when comparing acoustic speech metrics.

INTRODUCTION

As rates of Spanish speakers continue to rise in the United States, bilingualism is increasingly in demand, and the ability to speak two languages is an extremely helpful and valuable skill [1]. Differences in speech production between cultural and linguistic groups are equally as important. Many studies in the past have inspected the effect of ethnicity and language on speech fundamental frequency (fo). An investigation comparing the fo between English and Mandarin speakers found that in single-word utterances, Mandarin speakers had higher maximum and mean fo compared with English speakers. However, for a prose passage, the two languages were more similar, differing only in the mean fo, Mandarin again being higher [2]. In young children specifically, results show that in comparison to African American kindergartners, Hispanic kindergartners had an increased mean fo in spontaneous speech tasks [3]. In contrast to the growing literature that shows how languages differ in their fo profile, very few studies have compared voice and acoustic quality of bilingual speakers to monolingual speakers. In the limited current evidence, when analyzing the fo between bilinguals and monolinguals, results suggest a general trend of increased fo in bilingual speakers. One study inspected English/Russian and English/Cantonese speakers. They reported no significant differences in the English/Cantonese group, but a consistently higher mean fo in Russian than in English during spontaneous speech tasks [4].

In addition to language differences in fo between bilinguals and monolinguals, studies have also included other acoustic parameters to analyze differences in voice production associated with the language spoken. Some of these measures are jitter, shimmer, Harmonics-to-Noise Ratio (HNR) and Cepstral Peak Prominence Smoothed (CPPS). In the case of jitter and shimmer, measures of cycle-to-cycle perturbation of frequency and amplitude respectively, previous research in bilingual speakers is limited. However, in a comprehensive dissertation conducted at the University of Hong Kong, the author reports language and race effects in bilingual Mandarin/English speakers [5]. The study compares Chinese and American bilinguals who speak Mandarin and English, and findings report that higher jitter and shimmer are present in the majority language for each respective speaker-sample. Therefore, the Chinese bilingual speakers show higher jitter and shimmer in Mandarin than in English, while the American bilingual speakers showed higher jitter and shimmer in English. These findings confirm that a language effect is present in jitter and shimmer; however, this effect varies in speakers with different language competencies. Similarly, to jitter and shimmer, there is minimal evidence surrounding bilingual language differences in HNR. However, a study from the Massachusetts Institute of Technology that compares HNR in Japanese and English speakers shows a marked language difference, demonstrating higher HNR measurements in Japanese speakers when compared to English speakers [6]. In another study of HNR in bilingual Brazilian Portuguese/English speakers [7], variation in the mean HNR measurements in the bilingual speakers was reported. In the case of the CPPS, there is also scarce literature regarding the comparison of bilingual and monolingual speakers with one found study of various acoustic parameters in bilingual speakers of standard and local variants of the German language reported no significant difference in CPPS values [8].

Overall, there is minimal evidence that specifically focuses on the acoustic parameters of voice of bilingual speakers in comparison to monolinguals. Moreover, almost no evidence exists to explain acoustic differences between English-Spanish bilinguals and English monolinguals in areas of fundamental frequency, CPPS, jitter, shimmer, and HNR. Information needed to determine if, or to what degree, speaking a second language influences voice production. In order to address these issues, the current exploratory cross-sectional study of 40 monolingual English speakers and 13 bilingual English-Spanish speakers was designed with the aim to identify differences in voice acoustic parameters during English productions between monolingual and bilingual speakers.

METHODS

Design and Participants

This is an exploratory cross-sectional study with two groups of subjects: monolingual English speakers (n= 40), and bilingual English-Spanish speakers (n= 13). A higher proportion of female were present in both groups with a 55% of the monolingual speakers and 69% of the bilingual speakers being women. For sample size calculations, since acoustics parameters of voice are continuous variables, we used the formula for comparing two independent means. Assuming a difference between two means of 10 and an expected standard deviation of 10, it was determined that the study would require a sample size of 16 participants for each group (a total sample size of 32), to achieve a power of 80% and a level of significance of 0.05 (two sided) [9]. Inclusion criteria was normal hearing confirmed by hearing screening and reported no current voice problems the day of the experiment.

After approval of the Michigan State University Human Research Protection Program, participants were invited to take part of this study. Participants gave written informed consent to participate in this study, filled out a five-part questionnaire and recorded a voice sample.

All the participants were requested to fill in a self-administrated questionnaire where they indicated their native language. For this study, all the population reported that their native language was English. In the case of the bilingual group, they reported have studied at school or being exposed at home to Spanish as second language. During the initial interview with the bilingual English-Spanish researcher (LCCC), 5 out of 13 participants understood (listening ability) English better than Spanish, whereas eight participants understood both English and Spanish similarly well. In order to define the listening ability, the researcher started speaking in Spanish and if the participant was able to keep the conversation was classified with “good listening ability” in both languages, if the participant asked for repetition or clarification in English, he/she was classified with “better listening ability” in English than Spanish.

Data collection procedures

Questionnaire

The questionnaire presented to the participants consisted of 51 questions organized in five sections. Nevertheless, for this study, we used exclusively the information collected in the first part of the questionnaire. The first section included nine questions on socio-demographics (age, gender, ethnic category, and racial category), self-reported native language, history of speech/language therapy, and history of hearing or speech disorders.

Voice samples

All the participants were asked to read aloud a text in English consisting of the first six sentences of “The Rainbow Passage” [10], equal to about 30 seconds of speaking under a “virtual-simulated” acoustic condition of no artificial added background noise and medium Reverberation Time (0.8 sec). All the voice samples were recorded in a double walled sound isolation booth (2.5 × 2.75 m and h = 2 m) with a mid-frequency reverberation time of 0.05 seconds and a trend over the octave band almost flat. The background noise in the room was 25 dB(A). The setup of this study was used in previous experiments of our research group, and more details on the “virtual-simulated” environment can be found in previous publications [11, 12]. For this study, we analyzed the second sentence of the reading “The rainbow is a division of white light into many beautiful colors”.

Equipment

Voice samples were recorded using an omnidirectional microphone (M2211, NTi Audio, Tigards, OR, USA) placed at a fixed distance of 30 cm from the mouth of the participant. As in previous experiments [11, 12], the microphone output was split in two signals: one output was for direct recording and the other one for creating the virtual acoustic environment (no artificially added background noise with medium Reverberation Time). The first output (direct digital recording at 44,100 Hz) was performed using an external sound board (UH-7000 TASCAM, Teac Corporation, Montebello, CA, USA) connected to a laptop running Audacity 2.0.6 (SourceForge, La Jolla, CA). The second output (mix of the virtual acoustic environment and the participant’s voice) was digitally processed to add reverberation using a real-time effect processor (MX400, Lexicon, South Jordan, UT, USA) and played back to the participant using headphones (SRH840, Shure, Niles, IL, USA).

Voice parameters

Speech fundamental frequency (fo): The fundamental frequency of the voice is related to the average rate of vocal fold vibration. It is usually expressed on a continuum in units of cycles per second, Hertz [13], or on a musical scale in semitones [14].

Standard deviation of fundamental frequency (fo SD): The results related to fo may be shown as occurrence histograms, from which the standard deviation can be obtained. The standard deviation is a measure that is used to quantify the amount of variation or dispersion of a set of data values [15].

Jitter and Shimmer: These two parameters measure the cycle-to-cycle variations of fundamental frequency and amplitude, respectively, which have been largely, but not exclusively, used for the description of pathological voice quality [16]. Jitter specifically is commonly considered in dysphonic voice, as it is associated with a lack of control of the vocal folds [17]. Shimmer refers to amplitude variations in successive glottal impulses [18]. Shimmer is expressed in decibels as the variability of peak-to-peak amplitude. Jitter, as a voice parameter, represents pitch stability, and it is commonly used to inspect variations of the fundamental frequency in close cycles.

Harmonics-to-noise ratio (HNR): This parameter is a quantification of the periodic and aperiodic (noise) components of a voice signal, and specifically, HNR reflects the ratio between these components [19]. The calculation of HNR involves the acoustic waveform or a transformed representation of the waveform, and this calculation is estimated either on a time-domain basis or a frequency-domain basis [20]. HNR is frequently used to assess voice quality, and previous literature explains the correlation between HNR and the quantification and evaluation of hoarse voice quality [21].

Cepstral Peak Prominence Smoothed (CPPS): This parameter measures the degree of perturbation of a signal, and it describes the frequency distribution of the energy of the signal, the periodicity, and the harmonic components of the speech signal [22]. A normal voice with a well-defined harmonic structure will have a strong cepstral peak while an abnormal voice will lead to small CPPS due to relatively flat cepstral peak.

Data analysis

The second sentence of the rainbow passage was analyzed individually using custom MATLAB scripts, which preprocessed and managed the recorded samples. The MATLAB scripts used Praat (5.4.17) command line to extract the f_O contour (10 msec increments), Jitter, Shimmer, HNR, and CPPS. For the fo extraction, Praat options were based on the autocorrelation method. Specific parameters for the fo extraction, per PRAAT command line, were as follows: “To Pitch (ac)… 0.01 fo_min 15 yes 0.03 0.45 0.0025 0.35 0.20 fo_max”. The fo extraction thresholds (fo_min and fo_max) were set between 70 Hz and 450 Hz for females, and between 50 Hz and 350 Hz for males. Specific parameters for the HNR, Jitter and Shimmer extraction was on only the concatenated voice components of the speech as follows: “Voice report… 0 0 fo_min fo_max 1.3 1.6 0.03 0.4”. CPPS was obtained using a Praat function [23], which was modified to be controlled by MATLAB.

Considering the objective of this study, six dependent variables and one independent variable was defined. The dependent variables were continuous variables and included: fo, fo SD, jitter, shimmer, HNR and CPPS. The independent variables were dichotomous: second language knowledge (yes=bilinguals, no=monolinguals), and gender (male, female). Our dataset contained no missing values.

The statistical analysis was performed in three steps. First, normality of the data was assessed by means of the Shapiro-Wilk test. Second, differences in fo, fo SD, jitter, shimmer, HNR and CPPS between monolingual and bilingual speakers were assessed by Mann-Whitney U test. Third, Generalized Linear Models (GLM) were used to determine whether bilingualism was associated with differences in voice parameters (fo, fo SD, jitter, shimmer, HNR, CPPS). Because fo SD, HNR and CPPS were normally distributed, we used a GLM with a Normal distribution with these variables. On the contrary, fo (in Hz), Jitter and Shimmer were not normally distributed, therefore, we used a GLM with a Gamma distribution. The magnitude of the association was expressed by the beta (β) and its standard error (SE). All statistical analyses were performed by means of SPSS 21 (IBM Corporation).

RESULTS

Differences on fundamental frequency and standard deviation of fundamental frequency between monolingual English and bilingual English-Spanish speakers

Figures 1 and 2 shows the distribution of fundamental frequency calculated in the production of the second sentence of the rainbow passage. Fundamental frequency mean was similar in both group (bilingual and monolingual speakers). Nevertheless, distributions were significantly different (Mann Whitney U test −8.9, p-value 0.00 for females, and −2.7 p-value 0.01 for males). Analysis by gender shows that female monolingual speakers had higher fundamental frequency mode (210Hz vs 191Hz), and higher standard deviation of fundamental frequency (52Hz vs 45Hz) compared to female bilingual English-Spanish speakers. Male monolingual speakers also had higher fundamental frequency mode (109Hz vs 106Hz), but they had lower standard deviation of fundamental frequency (30Hz vs 32Hz) compared to male bilingual English-Spanish speakers. Figure 3 shows the distribution of standard deviation of fundamental frequency among bilinguals and monolinguals.

Figure 1.

Fundamental frequency female participants

Figure 2.

Fundamental frequency male participants

Figure 3.

Standard deviation of Fundamental frequency distribution among participants

Figure 3 shows the distribution of standard deviation of fundamental frequency among bilinguals and monolinguals.

Differences on jitter, shimmer and HNR differences between monolingual English speakers and bilingual English-Spanish speakers

Among male participants, bilingual speakers had higher jitter (Mean=3.0) and HNR (Mean=11.3) than monolingual speakers (jitter= 2.6; HNR= 10.4). On the contrary, female bilingual speakers had lower jitter (Mean=1.9) and shimmer (Mean=7.8) than monolingual speakers (jitter=2.2; shimmer= 9.2). Figures 4, 5 and 6 show the distribution of jitter, shimmer and harmonics-to-noise ratio among bilinguals and monolinguals.

Figure 4.

Jitter distribution among participants

Figure 5.

Shimmer distribution among participants

Figure 6.

Harmonics-to-noise ratio distribution among participants

Differences on CPPS between monolingual English speakers and bilingual English-Spanish speakers

As shown in Figure 7, the distribution of CPPS calculated in the production of the second sentence of the rainbow passage is similar in both groups. Mean value of CPPS among bilingual speakers was 12.6 dB and among monolingual speakers was 12.4 dB. Stratified analysis by gender shows that monolingual males had slightly higher CPPS compared with bilinguals (13.1 dB vs. 12.5 dB), whereas monolingual females had slightly lower CPPS compared with their bilingual pairs (11.9 dB vs. 12.6 dB). Nevertheless, none of these differences were statistically significant.

Figure 7.

Cepstral Peak Prominence Smoothed (CPPS) distribution among participants

Effect of language on fundamental frequency, standard deviation of fundamental frequency, jitter, shimmer, HNR and CPPS

Table 1 shows the results of the Generalized Linear Models. All the models were controlled by gender due to the differences in acoustic parameters related with this aspect. The results suggest that being a bilingual speaker was statistically associated with a slight decrease in shimmer (B= - 0.1) and a slight increase in HNR (B=1.1).

Table 1.

Associations between knowledge of second language (bilingualism) and acoustic parameters of voice

	Mean fo		fo SD		Jitter		Shimmer		HNR		CPPS
Variable	B	SE	B	SE	B	SE	B	SE	B	SE	B	SE
Intercept	4.8	0.0	26.1	2.4	1.0	0.1	0.1	0.0	10.4	0.3	12.9	0.4
Female	0.5*	0.0	19.0*	3.0	−0.2*	0.1	−0.1*	0.0	3.2*	0.4	−0.9	0.5
Bilingual speakers	0.0	0.0	−1.8	3.5	0.01	0.1	−0.1*	0.0	1.1*	0.5	0.3	0.6

B= Beta; SE= Standard Error

SD= Standard Deviation

^*= Statistically significant association

fo= Fundamental frequency

HNR= Harmonics-to-Noise Ratio

CPPS= Cepstral Peak Prominence Smoothed

DISCUSSION

The present research aimed to identify differences in voice acoustic parameters between monolingual and bilingual speakers. From the analysis of the data obtained from the participating speakers, three main results were found. First, there were statistically significant differences on the distribution of the fundamental frequency between monolingual and bilingual speakers. Second, male monolingual speakers had lower standard deviation of fundamental frequency compared to male bilingual English-Spanish speakers. Third, male bilingual speakers had higher jitter and HNR than monolingual speakers. On the contrary, female bilingual speakers had lower jitter and shimmer than monolingual speakers.

Concerning the first result, our findings show that monolingual participants had higher fundamental frequency mode compared with bilingual speakers (210Hz vs 191Hz for females; 109Hz vs 106Hz for males). This difference may be related with anatomical adjustments required to produce Spanish (as a second language). It has been reported that there is a language-dependency of the phonatory process, which implies taking into consideration several elements, such as anatomic and physiologic characteristics of the speaker and long-term muscular adjustments of the vocal apparatus, when analyzing voice production [24]. Therefore, bilingual speakers learn both the supra-laryngeal settings (articulation), and the laryngeal settings (phonation) of the second language (Spanish) [25]. In this way, a possible explanation for the lower fo mode among the bilingual speakers is the effect of the inclusion of Spanish sounds in their speech. Nevertheless, future studies with larger sample sizes are needed to confirm these results.

A second interesting results was that female bilingual speakers had a higher standard deviation of fundamental frequency when compared to monolingual speakers. The bilingual participants may have an increased spread of average fundamental frequency due to differences in laryngeal settings when making the phonatory switch from Spanish to English. In previous research that compares the fo and fo SD of Cantonese/English bilinguals and English-speaking monolinguals and found differences in fo when the bilinguals spoke Cantonese compared to English [26]. Their data attributed these findings to language differences, as Cantonese is a tonal language. Spanish may impact bilingual speakers who are making the phonatory switch between languages.

Concerning our third finding, male bilingual speakers had higher jitter and HNR than monolingual speakers. On the contrary, female bilingual speakers had lower jitter and shimmer than monolingual speakers. Previous research has compared the acoustic parameters of German, Italian, and Polish speakers, and found that language differences affect HNR, however, they did not report any significant language differences for jitter [27]. Perceptually, high HNR values are reported to affect the “brightness” of a voice, whereas low values may contribute to perceived “roughness.” In the case of the present study, bilinguals had higher measures of HNR, suggesting that language differences resulted in a brighter sounding voice when bilinguals made a phonatory switch to English.

Female bilinguals demonstrated a lower jitter, and shimmer compared to that of female monolingual speakers. Therefore, it seems likely that the female bilingual speakers adjusted their phonatory tract when speaking English as an attempt to have the same perceived pitch variation as monolingual English speakers. This variation supports the theory that language differences affect laryngeal settings when bilinguals switch from Spanish to English. Since there is an effect of gender on the acoustic parameters of the study, gender-specific thresholds would be useful in evaluations of bilingual speakers, providing a more accurate depiction of their acoustical measurements and acceptable ranges.

Overall, it is important to note that shimmer and HNR are also associated with being a bilingual. Specifically, it is these factors that differ between bilingual English-Spanish speakers and monolingual English speakers. When controlling for gender, results of the present study suggest that being a bilingual speaker was statistically associated with a slight decrease in shimmer (B= - 0.1) and a slight increase in HNR (B=1.1). Shimmer measures the variations in amplitude of successive glottal impulses [18], therefore it is an accurate representation of amplitude stability. In the current study, when compared to monolinguals, bilingual speakers had a statistically significant lower shimmer value of −0.1 dB. From this we can conclude that regardless of gender, bilingual English-Spanish speakers have a slightly more stable voice compared to monolinguals when speaking in English. This may be due to the phonatory switch between languages, in which Spanish sounds alter the production of fo and other voice parameters in English. In the current study, bilingual speakers had a significantly higher HNR of 1.1 dB when compared to monolingual speakers. This indicates that the bilingual speakers had increased periodicity. Again, this may be the result of phonatory switching in an attempt to alter voice parameters to sound proficient in English. While statistically significant, these differences may or may not be perceptually detectible.

As any research, this study has some limitations. First, we had a small sample size, which hampers the generalization of our results. Second, although 30 seconds of reading were recorded, we analyzed the second sentence of the reading “The rainbow is a division of white light into many beautiful colors”, which may be consider a short duration. Nevertheless, this sentence was selected for analysis since longer samples included more and varied pauses in speech which affected early iterations of the listening protocol. Using shorter samples (e.g. 2^nd sentence or 2^nd & 3^rd sentence of the Rainbow Passage) isn’t without precedent for both acoustic and perceptual analysis [28]. Therefore, the 2^nd sentence was used to reduce variability and increase comparability across samples. Future studies are recommended with bigger sample sizes and longer speech material.

In conclusion, the results from the present study suggest that language knowledge should be considered when measuring acoustical parameters regardless of gender. Future research should contribute to the understanding of acoustic characteristics in bilingual individuals by investigating changes in specific acoustic parameters, and how language knowledge impacts these changes.