From September 2013 to January 2014, we proposed to pregnant women attending the prenatal clinic of the Pitié-Salpêtrière University Hospital in Paris, France, to participate in a survey on maternal speech. They received oral and written information explaining that their participation would require filling out self-questionnaires, answering questions related to their emotional status, and being audiotaped when speaking to their future baby. The study was approved by the local Ethical Committee (CPPIDF6) under the number n°09012014. All participants gave a written consent. The inclusion criteria were mothers aged 18 or above, with a gestational age of 26–38 weeks, and able to understand the protocol. Indeed, during this period of pregnancy, future mothers are less concerned by the fetus viability. They begin preparing their meeting and relationship with the future child with more dreams and more detailed representations about the future baby (Ammaniti et al., 2000). Provided women were fluent in French, they could be included even if French was not their native language. Exclusion criteria were mental disorders and absence of health coverage according to the French ethical rules that require that studies be carried out only on people with health coverage. The mental disorder information were extracted from the obstetric record. Mental disorders were considered to be present if the pregnant woman was cared or treated for mental disorder by specialist before the pregnancy.

Statistical analyses were performed using R Software, Version 2.12.2. For all tests, the level of significance alpha was fixed at 5%. Given the sample size and the exploratory nature of the study, we used univariate analysis only. Quantitative variables were presented as the mean, standard deviation, and range. Qualitative variables were presented as frequencies.

First, we explored the correlation between maternal first perception of fetal movements and first vocalizations to the fetus.

We then successively estimated the relationship between:

The relationship between two continuous variables was either tested using Pearson r or Spearman rho, depending on the validity of the assumptions. The relationship between a continuous and a binary variable (fetal risk) was either tested using the Welch t-test or Wilcoxon rank sum test, depending on the validity of the assumptions. Finally, we estimated the agreement between our experts' measures and the algorithm's measures using ICC (single random raters, ICC2) and calculated the 95% confidence interval (R psych package).

One hundred forty-five pregnant women were considered eligible from September 2013 to January 2014. Thirty-five of them agreed to participate in the study and were enrolled. Four women participated in the clinical part of the study but did not record audio data; five audio records were not exploitable because of technical difficulties. Thus, 26 audio records were used for analysis. The most frequent reason declared by the invited mothers to decline the participation to the research was the lack of available time, considering that the interview needed at least 1 h and 30 min. Indeed in order to avoid displacements, we proposed the study to mothers who were already present at the Hospital for a pregnancy consultation.

Table 1 summarizes the study sample in terms of socio-demographics, life events and pregnancy risk factors, delivery, psychopathology and fetal-oriented interaction variables. Mothers presented a high number of life events and significant obstetrical (60%) and medical (65.7%) history: endocrinologic conditions (N = 7), multisystemic pathologies (N = 3), neurologic disorders (N = 4), psychiatric disorders (N = 2), uterine anomaly (N = 2), social precariousness (N = 2), and cardiac anomaly (N = 1). Regarding fetal risk, we found trisomy 21 risk (N = 3), intrauterine growth retardation (N = 2), premature delivery threats (N = 2), cardiovascular anomalies (N = 1), drug exposure (N = 1), and prior history of neonatal death (N = 1). This could be related to recruitment inside a free public university hospital in a maternity unit specializing in complex cases.

Nevertheless, our sample presented a low mean level of anxiety and depression scores and a high percentage of breastfeeding compared to the French general population (68.1–70.5% in the general population with 59% exclusive breastfeeding https://drees.solidarites-sante.gouv.fr/IMG/pdf/dt68-sources_et_methodes.pdf; Kersuzan et al., 2014).

Mothers declared to start speaking or vocalizing to their fetus on average at 3.63 (±1.64) months during pregnancy. Additionally, they started perceiving fetal movement on average at 2.83 (±1.89) months during pregnancy. We found a significant correlation between speaking to the fetus and perceiving fetal movements (Figure 2, right).

Regarding audio analyses, only 26 mothers were included because of technical issues (see Figure 1 flow chart). For low-level audio analysis (quantitative speech analysis), a total of 856 vocalization segments (mean vocalization number = 32.92) were detected. The duration of the vocalizations ranged from 0 to 3.95 s during the 3-min audio record (Table 2). The vocalization ratio (vocalization time during the 3 min) ranged from 0 to 65%. Indeed, two mothers did not speak during the 3-min recording.

Regarding high-level audio analysis (qualitative affective speech analysis), two complementary methods were performed: a qualitative manual annotation of maternal vocalization to the fetus by two expert psycholinguists and an automatic classification. In both clinical expert and automatized classifications, we found that pregnant women when speaking to their fetus (FDS) used sometimes a specific prosody that usually characterized motherese (or emotional IDS), which we called emotional fetal directed speech (e-FDS). We called the FDS without motherese characteristics “non-e-FDS.” The automatic classification yielded a mean e-FDS ratio during the 3 min of 0.12, whereas the expert classification found a mean e-FDS ratio during the 3 min of 0.14 (Table 2). Figure 2 (left) shows the strong and significant correlation between expert and automatic classification on e-FDS recognition. We also calculated the intraclass correlation (ICC) between the “two” raters (the expert and the algorithm) and found a good and very significant ICC (ICC = 0.79 (95% CI: 0.59–0.90), p < 0.001).

Given the limited sample size, we used only exploratory univariate analysis to address whether some stress or psychopathological variables could influence the ability to produce e-FDS. We found no association between speaking to the fetus (whether prosody had characteristics of e-FDS or not) and being a fetus at risk during pregnancy (correlation ratio = 0.36 (±0.2) and 0.33 (±0.2), respectively, t-test, p = 0.69).

Figure 3 shows the correlation between the speaking-to-fetus ratio and the Covy anxiety score, Raskin depression score and number of maternal stressful events. As shown, we found no correlation with the Covy anxiety score, a tendential negative correlation with the number of maternal stressful events, and a significant negative correlation with the Raskin depression score (ρ = −0.4, p = 0.046), meaning that the more the future mothers were depressed during pregnancy, the less they spoke to their fetuses during the experiment.

We performed the same analyses using only the e-FDS/vocalization ratio. None of the variables modulated the e-FDS/vocalization ratio, meaning that the impact of the Raskin depression score was on speaking to the fetus as a whole whether the pregnant women had e-FDS prosody or not. However, there was a trend toward a negative correlation between the e-FDS/vocalization ratio and the number of maternal stressful events (rho = −0.421, p = 0.072).

To answer this question, we proposed to address two different issues: (1) Is the mother speaking to the fetus during the experiment truly oriented toward the fetus? And (2) does FDS include some sequences that share the same prosodic characteristics of postnatal IDS (motherese)?

To address this first question (Is the mother speaking to the fetus during the experiment truly oriented toward the fetus?), we explored whether the pregnant women reported spontaneous moments of speaking and vocalization with their infant to come. For the mothers who reported doing so (n = 26), mothers started speaking or vocalizing with their fetus on average at 3.63 months during pregnancy (Table 1). Additionally, they started perceiving fetal movement on average at 2.83 months during pregnancy. This means that they could feel physically the existence of their fetus before they reported speaking to their fetus. Given the significant correlation between speaking to the fetus and perceiving fetal movement gestational ages (Figure 2, right), we can hypothesize that speaking to the fetus was indeed oriented toward the fetus. This result supports the hypothesis of a preliminary dialogue between future mothers and their fetus, as shown when mothers observed fetal movements during ultrasound scans that were interpreted by mothers as a response or solicitation from the fetus. Mirroring movements were seen as motor turn taking (Ammaniti et al., 2010). We believe that the current results on e-FDS are in the same vein and support the idea that prenatal development influences maternal infant attachment (Ammaniti et al., 2014; Feldman, 2016; Malm et al., 2016) and maternal representations of her future child (Viaux-Savelon et al., 2012, 2020).

Regarding the second question (do pregnant mothers sometimes use a motherese prosody (or here e-FDS) when speaking to their fetus?), our results show that futures mothers can use motherese prosody in their fetal-directed speech. The “manual” study of acoustic components of the voice takes a very long time and only allows the study of very short voice segments. The use of an automatic classifier allows extensive study of all vocalizations based on their acoustic characteristics and open perspectives for larger studies, and the machine learning classifier remains blind to the experiment or context. In this study, in addition to the expert “manual” categorization, the presence of e-FDS is confirmed by automatic measures that are strongly objective. Indeed, we found a strong and significant correlation between expert and automatic classification on e-FDS recognition (ρ = 0.87 p < 0.01) and a good and very significant ICC between expert and algorithm [ICC = 0.79 (95% CI: 0.59–0.90), p < 0.001]. This methodology of motherese detection using an algorithm has already shown robustness, as we have been able to distinguish motherese in early interaction with children with pathological outcome (Cohen et al., 2013), with both mothers and fathers (Weisman et al., 2016), and in five different languages (Parlato et al., 2020). Here, automatic annotation was useful to confirm that the prosody used during FDS shared the same characteristics of motherese. Manual and automatic labeling comparison realizes a validation of the two methods.

As expected, despite the limited sample size, the results show that the more the future mothers were depressed during pregnancy, the less they spoke to their fetuses during the experiment (Figure 3). We also found a tendency for a significant negative correlation between stressful life events and the speaking-to-fetus ratio (p = 0.69). These results are contingent with previous studies that have shown the impact of maternal prenatal states. Prenatal stress, particularly concerning prenatal diagnosis, increases the level of anxiety, disrupts the emotional investment of the parents toward the fetus (Watson et al., 2002; Petersen and Jahn, 2008; Kaasen et al., 2010) and disrupts parent-infant interactions after birth (Viaux-Savelon et al., 2012). In addition, pregnant women with depressive and anxiety symptoms talk and sing less to their fetuses (Hernandez-Reif et al., 2018).

Regarding fetal-directed speech quality (e-FDS or fetus-directed motherese), we found only a trend toward a negative correlation between the e-FDS/speaking to fetus ratio and the number of maternal stressful events. Thus, a high number of stressful events may reduce mothers' affective involvement with their future infant. We know that depressed mothers of young infants are not only less likely to speak with them (Herrera et al., 2004) but also more likely to display a reduced prosody of motherese with them (Bettes, 1988; Kaplan et al., 2001). Moreover, even when depressed mothers produce motherese, their infants fail to learn in response to their own-mother infant directed speech, despite normal competence (Kaplan et al., 1999, 2002). In our study with fetuses, we found that depressed mothers speak less to their fetus, but we did not find a correlation between depression score and e-FDS ratio. However, we cannot exclude that motherese quality could be poorer and less able to prepare language acquisition. As suggested in a recent study (Bartha-Doering et al., 2019), neural discrimination of speech could begin in utero. So we could expect that depression during the end of pregnancy may have repercussions on the first steps of language acquisition. However, given the small size of our sample and the exploratory nature of the study, we cannot conclude, and further studies with larger samples would be helpful.

Anxiety and depression status and a high level of stressful life events influence at least the quantity of fetal directed speech. Therefore, the quantity of fetal directed speech may be a sign to consider when detecting depression during pregnancy. Indeed, supporting these mothers in their investment toward the fetus and the future infant is compulsory for the prevention of later psychopathology (Mazzeschi et al., 2015; Røhder et al., 2020).

Finally, we found no significant association between speaking to the fetus (whether prosody had characteristics of e-FDS or not) and having a fetus at risk during pregnancy. This was not our hypothesis. However, the mothers' and fetuses' medical and obstetrical history of our population is very heterogeneous in this small sample size, and all risks may not be similar. In addition, the gestational age of the stressful event could also influence the impact on maternal representations and involvement. Again, a larger study would be necessary to better explore these factors with a comparison group according to the type of stress factor (e.g., mother complication/fetal complication/others) (Viaux-Savelon et al., 2012, 2020; Pisoni et al., 2016; Cuijlits et al., 2019).

As noted above, our sample was scarce (N = 26) and did not permit us to draw conclusions regarding the effects of various complex factors, such as maternal anxiety depression status or fetus risk. As many women declined participation in the study, we must discuss whether future mothers who agreed to participate could be more susceptible to speaking to their fetus than future mothers who declined participation. This study is only exploratory and used an experimental context. We also need to confirm that speaking to fetus also occurs spontaneously in more ecological contexts (e.g., at home). This might be achievable with automatic recording using portable devices for example. Additionally, we did not perform multivariate models to explore how relevant variables are robustly correlated or not. Nevertheless, it is important to note that one mother who declared before the audio recording she did not usually speak to her fetus actually spoke a lot to her during the recording. This may suggest that speaking to her fetus may be a widespread phenomenon.