Brain mechanism of unfamiliar and familiar voice processing: an activation likelihood estimation meta-analysis

Literature search and selection

In view of the wide range of aspects in the field of voice research, including both voice “identity research” and “voice emotion research”, the keywords chosen for this meta-analysis needed to target “VIP” relatively accurately. Considering that the research on “VIP” is often mixed with that on “voice emotion perception”, to avoid the potential influence of “emotion perception” on the coordinate “VIP” data, “perception” was omitted from the keywords after comprehensive consideration. Instead, “recognition” and “identification”, two relatively high-level identity processing terms, were selected to specify the research on “VIP” as much as possible. In addition, as mentioned in the Introduction section, the research on “VIP” has mixed terms, so our search results still contained quite a few studies on VIP.

It should be noted all the literature obtained by the search was screened according to VIP as defined in the introduction. In other words, the VIP studies in this study were all focused on the selective sensitivity of the human voice, that is, the lowest level of acoustic signal based on the perception of stranger identity. Studies that did not include voice identity processing, such as voice emotion processing, were excluded.

YuXiang Sun and XuePing Hu performed the search and identified relevant manuscripts, with no disagreements regarding manuscript eligibility. The literature search and selection were independently completed by the first author. Data extraction was independently completed by the first author. In the Web of Science database (voice recognition * OR voice identification * OR speaker recognition * OR speaker identification * OR talker recognition * OR talker identification *) AND (fMRI) were keywords searched for studies published before December 15, 2021, and a total of 378 studies were obtained. Through the advanced screening provided by Web of Science, 314 remaining papers were filtered after review papers, clinical trials, meeting, abstracts, case reports, unspecified, books and data papers were excluded.

According to the criterion of reading the title first and then the abstract, the literature was divided into three categories: VIP, VIR and VII. The screening criteria were as follows:

(1) Exclusion of studies on VIP, VIR and VII that were not defined in the introduction; (2) exclusion of speech studies; (3) exclusion of studies of nonauditory modalities; (4) exclusion of non-fMRI studies; (5) exclusion of studies based on ROI analysis or without reported coordinates; (6) exclusion of studies with abnormal populations; and (7) exclusion of studies with nonadult subjects.

A total of 17 available studies were obtained by screening, including five studies on VIP, seven studies on VIR, and five studies on VII. Although a total of 17 articles met the recommended minimum requirements for ALE analysis (Eickhoff et al., 2016) after exhaustively using as many keywords as possible for voice identity processing, a meta-analysis with a small sample size may not have had sufficient statistical power. Therefore, two schemes were carried out to complement the literature: the data of voice identity processing in multimodal identity processing that appeared in 314 studies were considered, and a total of five studies were obtained, including one study on VIP, one study on VIR, and three studies on VII. According to the previous practice of supplementing the literature (Parola et al., 2020), the studies that could be included in the meta-analysis were screened from the references of 17 eligible studies. This screening process ensured that all papers appeared in the Web of Science database and were voice identity processing studies. A total of 10 articles were obtained, including four studies on VIP, three studies on VIR, and 3 studies on VII. After careful reading, it was found that the keywords used in these studies did not have relevant pronouns related to voice identity processing but used other keywords to refer to them, such as “auditory perception” and “familiarity”.

In summary, a total of 32 articles were obtained by the retrieval and screening procedure, including 10 articles on VIP, 11 articles on VIR, and 11 articles on VII. A summary of the literature is shown in Table 2, and the process of literature retrieval and selection is shown in Fig. 1.

ALE meta-analysis

ALE is a coordinate-based meta-analysis method (CBMA) first proposed by Turkeltaub et al. (2002). Since then, researchers have continuously updated and iterated the algorithm to obtain higher statistical power (Eickhoff et al., 2012; Eickhoff et al., 2009; Turkeltaub et al., 2012). The basic logic of the ALE method is to model the activation focus as a 3D Gaussian distribution and then calculate the activation distribution map of the meta-analysis with each experiment as the unit. Specifically, to estimate the probability of cross-experimental activation of each voxel under certain conditions, the ALE method first calculates the “modelled activation” (MA) map of each experiment separately: the MA map of each experiment was calculated by combining the probability that a single voxel is included in each peak activation point of the experiment in the unit of voxel. Then, we took the union of these individual MA maps and calculated the ALE value based on the MA value to obtain the ALE value across the experiments. Finally, the possibility of activation across experiments was tested for significance. The null hypothesis of the ALE method was that there is no coincidence between the MA maps of each experiment in the meta-analysis, and all coincidence is caused by random factors. The ALE results calculated by the actual coordinates were compared with the ALE values of the null hypothesis for the significance test to obtain the results of the meta-analysis. For more detailed statistical principles of the ALE algorithm, please refer to Eickhoff et al. (2009), Eickhoff et al. (2012) and Turkeltaub et al. (2012).

According to the requirements of the meta-analysis, the fMRI coordinate data of activation peak points for voice identity processing in all 32 studies were manually exported and sorted into text documents. Then, the coordinate conversion function of GingerALE software was used to convert all coordinates reported in MNI space into Talairach space coordinates. The converted coordinates were then exported. After the coordinates were exported, the coordinate data were arranged into the coordinate data format required by GingerALE software, and then the ALE method was used for coordinate-based meta-analysis.

Four different ALE analyses were performed using GingerALE 3.0.2 (Eickhoff et al., 2009): an overall analysis of all the included voice identity processing studies, and an analysis of the combination of VIR and VII research were conducted to explore the brain activation of “familiar” voice identity processing. To investigate the brain activation of “unfamiliar” vocal identity processing by analysing studies on VIP and to explore the brain activation of familiar voice identity versus unfamiliar voice identity, results from analyses and were combined and compared. In the current study, VIR and VII results were combined and compared to those of VIP because the former two are more familiar voice identity processing than the latter, and the number of meta-analysis literatures can reach 22 after the combination of the former two, which exceeds the recommended minimum number of studies required by ALE analysis. According to the number of studies in each analysis, cluster FWE (Eickhoff et al., 2016) was used in ALE analysis for and , and the cluster distribution threshold (cluster-level FWE) was set as P < 0.01. The cluster formation threshold was set to P < 0.05 and 1,000 permutations. For the ALE analysis, because the research on VIP only included 10 articles, voxel FWE was adopted according to (Eickhoff et al., 2016), and the threshold was set as an uncorrected threshold of P < 0.05 and 1,000 permutations. Referring to the recommendations of Eickhoff et al. (2017), the ALE comparison and conjoint analysis of did not use FDR correction, but an uncorrected threshold of p < 0.05 and a minimum cluster standard of 200 mm3 were employed.

Voice identity processing

The ALE analysis results of 32 studies on voice identity processing (i.e., analysis ) are shown in Fig. 2A and Table 3. Significant activation was noted in two large clusters distributed on both sides of the brain in the temporal lobe, including the bilateral STG, the bilateral MTG, the left inferior temporal gyrus (ITG), and the right fusiform gyrus (FG), and in the frontal lobe, including the bilateral IFG, the right middle frontal gyrus (MFG), the left parahippocampal gyrus, the bilateral insula, the right claustrum and the right caudate nucleus.

Voice identity recognition/identification

For the ALE analysis of 22 VIR and VII (i.e., analysis ), significant activation was found in 3 clusters, including the temporal lobe (e.g., the bilateral STG, the bilateral MTG, the left ITG, and the right FG), the frontal lobe (e.g., the bilateral IFG and the right MFG), the left parahippocampal gyrus, the bilateral insula, and so on (for details, see Table 4 and Fig. 2B).

Voice identity perception

As shown in Fig. 2C and Table 5, the results of the ALE analysis of VIP revealed significant activation in the bilateral STG.

Comparison and conjoint analysis

The ALE analysis of VIR/I versus VIP, that is, (R + I) −P, showed significant activation in the right IFG and MFG (see Fig. 3A and Table 6). Conversely, the ALE analysis of VIP versus VIR/I, i.e., P−(R + I), showed significant activation in the bilateral STG, the bilateral MTG and the left IFG (see Fig. 3B and Table 6).

Moreover, through the conjoint ALE analysis of VIR/I and VIP, i.e., (R + I) + P, the results showed significantly distributed activation in the bilateral STG (see Fig. 3C and Table 6).

Brain mechanism of unfamiliar voice processing

As mentioned above, voice perception studies have repeatedly demonstrated the existence of voice-selective sensitive regions in the human temporal auditory cortex. Recently, Pernet et al. (2015) proposed a voice localizer scan to pinpoint the voice-sensitive areas of the temporal lobe, termed temporal voice areas (TVAs), which were mainly located in the bilateral STG/S. Of note, the TVA region is commonly associated with species-specific sensitivity of the human voice (Fecteau et al., 2004). In the present study, ALE analysis of VIP showed activation of the bilateral STG, which overlaps with TVA regions, suggesting that the bilateral STG is sensitive to human voice and responsible for processing acoustic information in voices. Combined with the concept of VIP we defined, that is, the listener only knows that the voice stimulus is the sound produced by the vibration of human vocal cords but does not know “who” it is. Thus, the voice identity is unfamiliar to the listener. In other words, the brain regions activated by the perception of a stranger’s voice can complete the basic acoustic processing of the unfamiliar voice input stimulus but cannot evoke the processing of the specific “identity” corresponding to the stimulus.

Of course, this could also be a result of the task or paradigm, since most voice perception experiments involve subjects passively listening to one stimulus after another (Fecteau et al., 2004; Fecteau et al., 2005; Warren et al., 2006); therefore, subjects do not have the opportunity to respond to the current stimulus and thus could not further compare the current input stimulus with various “voice identities” stored in their memory. In contrast, Roswandowitz, Swanborough & Fruehholz (2021) added the task of actively judging whether the current stimulus was a human voice and found that the IFG still had functional involvement when the voice was actively processed. In addition, activation of the IFG was found in two recent studies of VIP with large samples (Aglieri et al., 2018; Pernet et al., 2015), different from the active task of Roswandowitz, Swanborough & Fruehholz (2021). These two large-sample studies both reported activation of the inferior frontal gyrus under the condition of passive listening. They concluded that due to considerable interindividual variability in frontal lobe anatomy and function, frontal lobe activation was understandable in some subjects under large sample conditions. At the same time, voice is closely related to speech and mixed with certain language content, so part of the frontal lobe can be used as the voice-sensitive extension area outside the temporal lobe.

Therefore, the current ALE analysis results of VIP should be interpreted with caution given the passive listening design adopted in most experiments, considering the difference in subject sample size, or based on whether the stimulus materials contain language information.

Brain mechanism of voice processing in familiar people

The activation of the superior temporal cortex (STC) in the transverse temporal gyrus extending to the superior temporal gyrus was consistent with previous studies, and this region was also consistent with the classic TVA region (Belin et al., 2000; Pernet et al., 2015), suggesting that this region activates and participates in acoustic processing of a voice regardless of voice familiarity (Andics et al., 2010; Belin et al., 2011). The activation of the MTG is consistent with the findings of Bethmann & Brechmann (2014) and Bethmann, Scheich & Brechmann (2012), who assessed the degree of voice-specific processing in the superior/middle temporal cortex and found that the difference between the specificity of voice processing and other sounds increased from the superior temporal gyrus to the middle temporal gyrus. In the current study, activation of the MTG may expand the TVA region in the classical temporal lobe, suggesting that the vocal selective brain region includes the area from the STC to the MTG, which is responsible for the basic acoustic processing of input stimuli.

Moreover, the ITG has been recruited to process the identity of familiar speakers. A previous study found that a patient with bilateral temporal lobe atrophy and more obvious atrophy of the right ITG (including the fusiform gyrus) had severe defects in familiar voice identification (Roswandowitz, Maguinness & Von Kriegstein, 2018). As the homologous contralateral brain region of the right ITG, the role of the left ITG in the identification process of familiar voice needs further investigation. In addition, as another possible explanation for the activation of the left ITG, it has been found that the ITG is activated during tone tasks (Stevens, 2004), and a voice contains a large amount of spectral information (Lattner, Meyer & Friederici, 2005). The left ITG may participate in the spectral information processing of a voice together with the STC.

Consistent with previous findings on high-familiarity voice identity processing (Hoelig et al., 2017; Joassin, Maurage & Campanella, 2011; Shah et al., 2001; Von Kriegstein et al., 2005), the activation in the right FG suggested that it may be involved in ascertaining a highly familiar voice identity, especially when the information of the voice identity has multimodal semantic associations. As Von Kriegstein et al. (2005) found, when subjects recognized the voice identity of a completely familiar person, there was functional interaction between the face recognition area and the voice-sensitive area, indicating that listeners might automatically retrieve the personal identity corresponding to the voice. In addition, it was found that in the process of familiar voice identification in early blind patients, there was functional connectivity between the voice selection area of the left temporal lobe and the right FG according to Dormal et al. (2017). Moreover, the activation of the FG was affected by the consistency of voice identity and showed more activity during the recognition of familiar voices for advanced or congenital blindness (Hoelig et al., 2014a; Hoelig et al., 2014b). Evidence from both sighted and blind people suggests an integral role for the fusiform gyrus in the process of identity recognition/identification of familiar voices.

Importantly, extensive frontal lobe regions (e.g., the bilateral IFG and the right MFG) were involved in high-familiarity voice identity processing, which is consistent with these previous studies (Aglieri et al., 2021; Andics, McQueen & Petersson, 2013; Blank, Wieland & Von Kriegstein, 2014; Latinus, Crabbe & Belin, 2011; Rama et al., 2004; Relander & Rama, 2009; Zaeske, Hasan & Belin, 2017), when the voice stimuli were trained such that they reached a degree of familiarity. In addition, studies of voice prototypes have found that bilateral IFG activation is often associated with acoustic information processing of a voice (Andics et al., 2010). Furthermore, the activation of the frontal lobe may be a function of cognitive control in voice identity, and the IFG and MFG are highly correlated with cognitive control functions. This finding may indicate that subjects need to inhibit the interference information, then employ task-related resources to retrieve personal information in the memory system and compare it with the voice currently being heard, and finally complete voice identity recognition and identification (Hu et al., 2017).

The hippocampus is commonly associated with memory function (Schacter et al., 1996; Stark & Squire, 2001) and is also associated with participants’ familiarity with voice identity (Shah et al., 2001) or retention of working memory for voice identity (Rama & Courtney, 2005). Therefore, the hippocampus may be related to the memory content corresponding to the retrieval of a familiar voice identity. In addition, it has been found that the bilateral insula and the bilateral subfrontal cortex are activated in the processing of acoustic information of a voice (Andics et al., 2010; Latinus, Crabbe & Belin, 2011). Pertinently, the insula is also associated with working memory and plays an important role in maintaining and recognizing the identity of familiar speakers’ voices (Rama et al., 2004; Relander & Rama, 2009). It is worth noting that the activation of the claustrum and caudate nucleus has never been reported in previous studies. Structurally, the claustrum is adjacent to the lentiform nucleus, but its function is unknown. The caudate nucleus is closely connected to the lentiform nucleus, and both are part of the striatum, which in vertebrates is mainly involved in the execution of fine movements (Grillner et al., 2005). In this study, we found that the activation of the claustrum and caudate nucleus in voice processing, one possibility may be the result of the concomitant activation of the insula, which may not be associated with voice processing itself.

Together, the ALE analysis results showed that VIR/I involves a wide range of brain networks, including basic acoustic analysis, multimodal information retrieval, cognitive regulation, memory retrieval and other functions.

Similarities and differences in voice identity processing mechanisms between familiar speakers and unfamiliar speakers

Limitations

As mentioned in the introduction, the recent research on voice identity processing has mixed terminology. At the same time, the field of VIP intersects with the field of voice emotion perception, so the data related to identify processing may be affected by other factors, which then leads to considerable complications in literature retrieval and coordinate data sorting.

Although this study encompassed the relevant literature as much as possible, after identifying the relevant research (Eickhoff et al., 2016), it was found that the number of studies corresponding to the three subcategories, which included VIP, VIR and VII, was still insufficient to meet the requirements of ALE analysis based on cluster-FWE correction alone. Therefore, it was impossible to compare the three categories separately; otherwise, results with very low statistical power would be obtained, which is a weakness of this study. We are not sure whether there is statistical bias in this method, but we have improved the statistical efficiency in ALE analysis as much as possible. It is expected that a large number of well-defined studies and data will be used to supplement and improve upon this study in the future; in addition, we will continue to assess relevant studies. At the same time, we will further expand the search scope and sort out the relevant literature in these three categories, with the goal of further improving on the present study. In the future, we hope to conduct a series of experiments to compare the brain activation of voice identities with different degrees of familiarity (perception, recognition and identification) to further verify the results of this meta-analysis.