On the neural underpinnings of voice discrimination, recognition, and identification

Electrophysiological observations.

In considering studies that use ERPs, the following brief review sets aside a body of work relating to the interplay of visual and vocal information in voice processing, which entails varying methodologies [for summaries of this work, see 41–44]. Early studies involving ERPs focused on the discrimination of human voices and non-human sounds (e.g., animal cries, bell sounds, tones, etc.), which showed distinct responses to voices with an onset as late as 400 ms or the N400 [45–47]. More recent reports have revealed that the discrimination of human voices compared to generic sounds is represented by early components, around 150 ms, which have come to be termed the “fronto-temporal positivity to voice” [FTPV; 48–51]. Thus, there is evidence that the earliest neural components that relate specifically to human voices are in the order of 150 ms post-onset. Given these results, one can logically assume that any processing of voice identity information would entail later-occurring ERPs as could be revealed on stimuli of IFVs.

Few studies, however, have investigated the processing of IFVs such as the voice of a close friend or family member. One exception is a study by Beauchemin, De Beaumont [3]. That ERP study focused on responses of listeners to IFVs (close relatives or long-time friends) using an auditory oddball paradigm in reference to the MMN components [52]. Short speech samples consisting of single vowels lasting some 200 ms produced by familiar and unfamiliar speakers (IFVs and UVs respectively) were presented in conditions of passive listening. The results showed distinct responses across IFVs and UVs peaking at 200 ms post-onset, in the MMN range. Similar results of MMNs have also been reported in studies involving newborns, suggesting that the ability to recognize voices arises in early development [53–55].

In a different study that also involved an auditory oddball paradigm, Graux, Gomot [56] compared the ERP of three sets of presented voices, including FVs, UVs, and participants’ own voices (designated as “self”). The results displayed a significant MMN between 180 and 210 ms post-onset (for FVs compared to UVs) and a significant difference on the P3a between 230 and 320 ms for FVs compared to self-voice. These results confirmed a previous report of a distinct process between self and familiar voices [4]. On the other hand, given that externally generated familiar voices are never heard as self-generated voices, it is difficult to extrapolate results on self voices to processes of voice recognition and identification [and see 57–60]. On their side, Holeckova, Fischer [61], after exposing participants to their own name pronounced by intimately familiar and unfamiliar speakers, reported a small effect on the P3, between 300 and 380 ms, but mostly on later-occuring ERP between 625 and 800.

Other studies have investigated ERPs to IFVs but with quite different results bearing specifically on voice identification. Of particular interest is a study by Schweinberger, Walther [62] based on a 2AFC task involving paired stimuli of two IFVs that were morphed to varying degrees with one another. ERPs in this paradigm reflected changes in voice identification when increasing the proportion of one IFV in the stimuli relative to another IFV. The experiment also included congruent/incongruent speech contexts with /aba/ and /iɡi/ serving to examine the effects of verbal contexts on responses. Importantly, the results showed two responses occurring at different time intervals. A first response to IFVs occurred at parietal sites starting at 250 ms post-onset when speech contexts were congruent whereas, when speech contexts were incongruent, a speech-independent response to IFVs appearing, not as short-time components, but as protracted changes between 300 ms and 600 ms post-onset, in the P3 range. It is useful to note that the authors used the designation “voice identification” in commenting on their results (while also expressing reservations on their interpretations owing to the small number of participants).

Another investigation that involved ERPs and more or less FVs was that of Gonzalez, Bobes Leon [63]. Their experiment used a go/no-go task with presented FVs and UVs in the context of a short phrase (the Spanish word /ola/). In this case, “FVs” referred to participants’ friends or colleagues so that it is unclear whether the stimuli could qualify as IFVs. The results showed ERP differences between FVs and UVs appearing between 280 and 840 ms post-onset, including a N250r and a P3, but again reflected protracted responses rather than short-time components as in Schweinberger, Walther [62].

Finally, one should note that it is often assumed in voice research that pitch (given by F₀) is voice-specific whereas such aspects can relate to speech processing as in the case of “tone languages” where pitch changes serve to distinguish between words. In a study involving ERPs (and fMRI), Zhang, Pugh [64] examined the varying responses obtained when listeners attend to changing lexical tones in two Cantonese words /ji/(produced with high or rising tones) and when they attend to changing voices (UVs of a male and female speaker producing the words). The design aimed to compare ERPs of talker and speech deviants with reference to a standard. The analyses of designated components showed talker-specific changes in P2, P3a, and on frontal negativities examined over an interval of 500–800 ms (parietal late components also appeared but were not analyzed). An important methodological implication of this study is that it showed task-dependent interactions between talker and speech processing where pitch could not be taken a priori as a property of “voices”. Moreover, the authors specified that the differences in F₀s between the male and female voices (101 Hz) exceeded differences in F₀s of speech contexts (56 Hz). There is much behavioral evidence that salient differences in voices can influence memory as opposed to less distinctive voices and such differences on distinctiveness are likely to reflect in ERPs. However, few reports specify F₀ values of voice stimuli, which may underlie the discrepancy in reported components of voice processing. But perhaps a more important source of variation is the length of the stimuli used across studies.

Generally, and in comparing various reports listed in Table 1, ERP responses to IFVs appear to involve short-time components between 150 and 320 ms but also prolonged responses with latencies up to 840 ms that have not been identified in terms of specific components. Although several methodological factors may underlie the discrepancies in reported latencies, one basic factor appears to be the duration of the stimuli, as seen in Table 1. In terms of reports involving IFVs, the stimuli length in studies by Schweinberger, Walther [62] and Gonzalez, Bobes Leon [63] provided sufficient dynamic spectral information so as permit speaker identification, whereas it can be questioned whether single vowels offer sufficient sensory information for this process [see 6].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Summary of ERP studies of voice processing arranged by type of stimuli and types of voices–intimately familiar voices (IFV), famous/familiar voices (FV), trained-to-familiar voices (TV) or unfamiliar/unknown voices (UV).

Only time windows in relation to voice processing are reported in the table.

https://doi.org/10.1371/journal.pone.0250214.t001

As for investigations that focus on stimuli classed as TVs and “famous” FVs, these stimuli involve, respectively, UVs that become familiar during a training phase of an experiment, or FVs from celebrities. Importantly, an experiment by Schweinberger [65] using a priming paradigm established that priming voices before the presentation of FVs or UVs creates a response at 200 ms post-onset indicating a voice-recognition response. However, a speaker-identity response for famous FVs was only observed in a time window between 450–800 ms (although the author did not label these sustained potentials identification responses). Contrasting with these results, several reports using TVs have not revealed responses in windows beyond 450 ms. Thus, the MEG study of Schall, Kiebel [69], based on TVs, used long sentence-length stimuli. After learning six voices with corresponding names, participants were asked to indicate if a speech sample and a name were matching or not. Significant responses to speaker identity were observed at 200 ms post-onset. Zäske, Volberg [70] similarly reported a significant difference in ERPs using an old/new task with TVs and long stimuli. TVs that were correctly identified elicited a greater positivity than UVs starting at 300 ms post-onset, although how this reflected a speaker-identity response was unclear since the responses occurred on identical linguistic stimuli (i.e., it was unclear whether identity information was processed independently of verbal contexts). A following study reported in Humble, Schweinberger [66] reported a similar old/new effect bearing on speaker identity, but this effect was observed later (500–800 ms) and was elicited following the presentation of stimuli different at learning and at test. Spreckelmeyer, Kutas [71] also reported a voice recognition response at around 300 ms post-onset during a same/different task involving pairs of UVs. Consistent with these results, Föcker, Hölig [67] reported rising negativity starting at 270 ms post-onset for person-incongruent dyads of TVs compared to person-congruent ones. Yet, in a very similar study, Föcker, Best [68] found a significant response to paired TVs in time windows between 200 to 250 ms and 350 to 550 ms. However, with the exception of Schweinberger [65], it is unclear how the paradigms in the preceding reports serve to distinguish responses bearing on a processing of speaker-identity information from those that reflect a recognition of voices. In fact, in many of the reports the terms voice recognition and voice or speaker identification are used interchangeably or with vague definitions.

Overall, neural responses that relate to the recognition of TVs appear to occur in the range between 200 and 370 ms post-onset (see Table 1). The experiments of Schweinberger [65], using FVs, yielded much later responses that could be related to speaker identification. This also applies to the report by Gonzalez, Bobes Leon [63]. In comparing these studies to others in Table 1, one notices that the reported long latencies ranging from about 500 to 840 ms post-onset appear for speech contexts consisting of at least a few syllables. As emphasized by many authors, stimuli length affects speaker recognition and identification in that longer stimuli generally entail greater phonetic variability and spectro-temporal information [6, 7, 12, 13] (see also [72] and [73] for further discussion on this topic). In understanding the differences between responses at long latencies and those that occur at about 200–370 ms, it should also be weighed that stimuli of famous FVs can associate to varying degrees with a multimodal episodic memory of speakers, whereas TVs, which are experienced in a laboratory setting or through repeated audio presentations, may not serve to constitute such multimodal representations. This is not an issue when using IFVs where sensory experiences spanning years associates with the voice of an individual. Such differences could well underlie the separate responses across 200–370 ms and 500–840 ms where the first response reflects voice recognition and a later-occurring response may reflect a processing of identity information that bears on episodic or semantic memory of a speaker. However, it remains unclear whether this is actually the case given that, except for Schweinberger [65], studies have not compared responses to TVs and IFVs. In interpreting the time windows reported in Table 1, it is interesting to note that Schweinberger, Walther [62] is the only study where voice identification was associated with ERP response between 250 and 600 ms post-onset. As noted, the two other studies where voice identification possibly occurred, Schweinberger [65] and Gonzalez, Bobes Leon [63], showed responses ranging from about 500 ms to 840 ms. One potential explanation for earlier response times reported by Schweinberger, Walther [62] is that all presented voices were IFVs although participants did not specify if they knew the speaker and were aware that any of the voices they would hear was and IFV. This accurate prediction could have facilitated the identification process and therefore fasten the EEG response.

The present study

In terms of the above research, one can surmise that EEG/MEG investigations of voice processing have not circumscribed the time course of fundamentally different processes of voice recognition and voice identification. Moreover, as summarized in Table 1, few studies focus on IFVs using sufficiently long speech samples that support accurate speaker identification [6]. Of the studies that do use stimuli consisting of at least a few syllables, separate responses appear on different time windows. Thus, while IFVs elicit responses in a 150–320 ms window, they also associate with prolonged responses as late as 840 ms post voice onset. The above discussion suggests that one reason for these prolonged responses is that IFVs and FVs carry information that links to semantic memory of a speaker such that the late responses reflect a process of voice identification.

The present study aims to bring further evidence supporting this latter view by examining the following prediction. Specifically, it is hypothesized that IFVs, compared to TVs and UVs, elicit voice recognition responses in a window of 150–320 ms, in the range of the P2 ERP component, as well as later-occurring responses extending beyond 450 ms, encompassing slow ERP waves, suggesting a distinct process of identification. This prediction also serves to clarify the effects of different types of voice stimuli, which are often indiscriminately associated with recognition and/or identification. Studies have frequently suggested similar responses for known voices regardless of whether these are IFVs, FVs, or TVs (as outlined in Table 1). Yet, as Kreiman and Sidtis [39] note, IFVs are distinctly processed, which should reflect in differential neural responses. It should be noted, however, that reports confirming these differential responses point to changes over long time frames (roughly 500–840 ms) and not to particular short-time ERP component [as in 63]. Indeed, studies of responses to IFVs that refer to components such as MMNs and FTPVs have used brief stimuli like single syllables which, as noted, may not provide sufficient information for processing voice identity [cf. 3]. For this reason, the present research is not driven by an assumption of particular components but instead explores how IFVs, TVs, and UVs elicit differential electric brain responses reflecting distinct processes of voice recognition and identification.

Participants

Thirteen participants (8 females), aged between 21 and 43 years (mean = 30.81, s.d. = 5.14) completed the study. They were all native speakers of Quebec French except one speaker who learned Quebec French at four years of age. All were dominant right handers according to a standard questionnaire [74] and had normal hearing as established by an audiometric screening test. A forward and backward digit-span test [WAIS-III, 75] confirmed normal memory performance for all participants. It should be noted that participants recruited in the present study were selected following the recommendation of a member of an original pool of 36 male volunteers from whom voice samples were recorded and analyzed to create the stimuli in the present study [2]. Each volunteer in this pool provided the names of a family members, close friends, or life partners. The recruitment of participants in the present study was limited to these named individuals who could be matched to one target IFV in a set of otherwise unfamiliar voices but where all voices reflected speakers with similar Speaking Fundamental Frequency (SF0) to within one semitone, as described subsequently. The “intimate familiarity” of a target IFV was established via a questionnaire and criteria that were elaborated in a previous behavioural study [2]. The fact that participants were selected by reference to an IFV which had to be similar to other unfamiliar voices in a set restricted the recruitment to a small number of specific individuals (i.e., participants that could be matched to one IFV in a set of highly similar voices). All participants were paid, and written informed consent was obtained following guidelines of the Ethics Committee of CIUSS du Nord-de-l’île-de-Montréal at Rivière-des-Prairies Hospital (Montreal, QC) which also approved the present research.

Stimuli

The voice stimuli were eight four-syllable utterances reflecting usual speech and expressions. These utterances, listed in Table 2, were produced by 12 native speakers of Quebec French who, as noted above, all had similar SF0 and had no discernible regional accents. The length of the stimuli (4 syllables) was decided following the results of Plante-Hébert and Boucher [6] and other observations relating to the length of contexts required for accurate speaker identification [7, 12, 13, 22, 72, 76]. These studies, especially the one from Plante-Hébert and Boucher, refer to stimuli exceeding one syllable for correct identification. Average SF0 was controlled and similar across the voice stimuli used in the experiment (the stimuli were spoken with a neutral intonation and cross-speaker differences in SFO for the voice samples did not exceed one semitone).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. The four-syllable utterances used as voice stimuli.

Transcripts in regular orthographic Quebec French and IPA.

https://doi.org/10.1371/journal.pone.0250214.t002

Finally, as indicated in Table 2, each utterance contained a number of nasal sounds, which have been shown to facilitate speaker identification, likely because they provide additional information on speaker physiology in relation to resonance cavities [5, 8–11, 77]. The voice stimuli were produced in a conversational fashion at steady rates and recorded in a sound-treated booth using an omnidirectional headset microphone (C477 WRL, AKG) and a 16-bit external sound card set to a sampling rate of 44,1 kHz (Fast-track Ultra, M-Audio). While recording these stimuli, the speakers produced each utterance after listening to an audio pacer consisting of separate tones. This ensured the production of similar timing and prosody across utterances. The recorded signals were amplitude normalized and each stimulus was segmented as a separate audio file. The onsets of the speech signals in the audio files were aligned so that the perceptual-center (P-center) of the first syllable of all utterances was at 200 ms from the beginning of the file. Alignment in terms of P-centers [described in 78, 79] insures that perceptual onsets of speech stimuli are stable and reduces jitter in EEG responses at the onset [see 80]. The overall length of the signals ranged from 618 ms to 1085 ms (mean of 818 ms, SD of 83 ms). Overall, the stimuli used in the present experiment respected generally admitted guidelines for the elaboration of voice line-ups in forensic applications [81–84].

Pre-test stimuli validation

As a preliminary verification of the stimuli used in the present study, we conducted a pretest involving four volunteers that did not know any of the presented voices. The purpose was to establish whether equal numbers of presentations of the different voices and utterance contexts created non-specific ERPs. The test conditions were the same as during the experiment described below and each volunteer was exposed to a total of 10 trials per voice per utterance. The pretest confirmed that, in presenting different voices an identical number of times, average ERPs did not visually differ across utterance contexts. However, one of the voices had to be removed due to an unexplained difference in ERPs compared to the other voices. The pretest also confirmed that the multiple presentations of the different utterances did not have an effect on average ERPs across voices. In sum, variations in ERPs under the present test conditions can be related specifically to familiarity and frequency of presentation rather than utterance contexts or vocal idiosyncrasies.

Procedure

Audio files containing the stimuli were arranged in eight blocks, each reflecting a specific utterance of Table 2. Within each block, the voices were randomized with the restriction that no consecutive presentation contained the same voice. Of the eight blocks of stimuli, four served to record EEG responses, and these alternated with four blocks that served to collect behavioural responses on speaker identity. Specifically, the EEG-recording blocks were ordered such that the first, third, fifth, and seventh blocks each contained 240 trials of passive listening. The four other alternating blocks each contained 60 trials where listeners identified the IFV using a key press. The latter blocks of trials were reduced in number so as to limit the overall test duration while allowing to collect behavioral confirmation of IFV identification. All blocks bore presentations of different types of voices in varying proportions: a frequently presented IFV (33.33% of trials), a frequently presented TV (33.33% of trials), and twelve rarely presented UVs (each UV was presented on 2.77% of trials). Note that the 13 participants were recruited on the basis that they were intimately familiar with only one target voice (IFV) in the presented stimuli. Thus, 12 different voices were presented in the blocks but only one voice was intimately familiar to one participant.

Participants listened to the utterance stimuli via insert earphones (E-A-Rtone 3A, EAR Auditory Systems) and the amplitude of the audio signal was calibrated so as to obtain peak levels of 74 dBa at the inserts. The stimuli were played back using E-prime 1.0 (Psychology Software Tools). Trials were separated by an inter-stimulus interval (ISI) that varied randomly from 500 ms to 650 ms in steps of 50 ms to minimize anticipation effects. In listening to the stimuli, the participants were sitting at a distance of 180 cm from a blank computer screen with a fixation cross. They were asked to listen to the stimuli and keep their eyes on the fixation cross. For the four behavioural blocks, participants were also required to keep the fingers of their dominant hand positioned on a mouse and to indicate as quickly as possible if the voice heard during each trial was the familiar one or unfamiliar by pressing either the left or the right mouse key, respectively (this was reversed for half of the participants).

EEG recordings and analyses

EEG signals were recorded throughout the experiment (including behavioural blocks that were not included in the EEG analyses). The recordings were performed according to the international 10–20 system and with an ASA-lab EEG/ERP 64 channels amplifier (ANT neuro). An online average reference was used and signals were digitized at sampling rate of 1000 Hz. Eye movements and blinks were recorded using four electrodes placed above and below the dominant eye (VEOG) and at the outer canthus of each eye (HEOG). AFz was used as ground and all other 64 channels were kept below 10 kΩ impedance during the recordings.

Offline, the recordings were band-pass filtered (0.3–30 Hz) and blinks were removed using ASA software (ANT neuro). All other artefacts exceeding a standard deviation of 20 μV within a sliding window of 200 ms were automatically removed with Eeprobe GUI (version 1.2.0.2, ANT Software). EEG recordings were then averaged across blocks and by types of voices (IFV, TV, and UV) using Fieldtrip [85], an open-source toolbox for MatLab (R2017b 9.3). Each trial in the recordings was epoched between 200 ms before and 1000 ms after each stimulus onset and the 200 ms pre-stimulus interval was used for baseline correction.

Visual inspection of the averaged signal for all conditions allowed to easily identify a P1-N1-P2 complex, directly followed by a negative deflection between 300 and 350 ms post stimulus onset, in the range of the N250, and a late positive component (LPC) extending to the end of the analysis window. Considering the data in Table 1, the P2 peak on right centro-frontal sites, the N250 peak on left fronto-central sites and the LPC on both right centro-frontal sites and left/middle centro parietal sites were of particular interest in the present study.

The P2 was peaking on frontal sites between 200 ms and 250 ms post stimuli onset. An ANOVA with three within subjects factors was carried out on the mean amplitudes between 200 and 250 ms. The factors included were voice condition (IFVs, TVs and UVs), site (F, FC, C and CP) and laterality (right and left hemispheres)

The N250 and the slow waves ERPs had a wider scope than the P2. In order to reduce statistical analyses for those two components, pools of electrodes were created to represent six scalp regions. The regions included the following electrodes and will be referred to as: middle centro-frontal (MCF; Fz, F1, F2, FCz, FC1, FC2, Cz, C1, C2), right centro-frontal (RCF; F4, F6, FC4, FC6, C4, C6), left centro-frontal (LCF; F3, F5, FC3, FC5, C3, C5), middle centro-parietal (MCP; CPz, Pz, POz, CP1, CP2, P1, P2), right centro-parietal (RCP; CP4, CP6, P4, P6, PO4, PO6) and left centro-parietal (LCP; CP3, CP5, P3, P5, PO3, PO5). Outlining electrodes were not included since ERP of interest were not elicited on those sites.

Statistical analyses were carried out using 50 ms mean amplitude samples to compare the time-course of ERP activity between experimental conditions (IFV, TV, UV) on the N250 and de slow waves ERPs. The analyses window for the N250 was between 300 ms and 350 ms while successive 50 ms mean amplitude samples were used between 450 ms and 850 ms post-stimulus onset to investigate longer slow waves ERPs. The mean amplitudes were calculated using MatLab (R2017b 9.3). Repeated measures analyses of variance (ANOVA) were then carried out using the open-source software JASP (version 0.13) with two within-subjects factors: voice condition (IFV, TV and UV) and scalp region (6 levels: MCF, RCF, LCF, MCP, RCP and LCP). Huynh-Feldt correction was applied if required and the alpha level was set at p < 0.05.

Behavioural data

On the analyses of behavioural responses, all responses exceeding 1300 ms were excluded (22.30%). For the remaining trials, the overall accuracy of identification of the IFV was 98.18%. The false alarm rate, that is when either TV or UV were falsely identified as an IFV, was 0.35%. The misses, or when IFV was designated as UV, represented 1.82% of response. Most of the time, when participants made mistakes, they spontaneously informed the experimenter that they were aware of their error. These results establish that the voice stimuli were readily identified by participants.

The P2

For the P2. the three levels repeated measures ANOVA carried out on mean amplitudes between 200 ms and 250 ms reached significance for the main effect of site, F(1.883,22.595) = 17.348, p < .001, η² = .297 as well as for the interaction between laterality x voice conditions, F(2,24) = 3.868, p = .035, η² = .01. Planned comparison with Bonferroni-corrected t-tests for the voice conditions on each sites revealed significant differences between IFV and TV on F4, t(12) = -2.3, p = .04 d = .638 and between IFV and UV on FC4, t(12) = 2.498, p = .028, d = .693, C4. t(12) = -2.25, p = .044, d = .624 and CP4, t(12) = -2.716, p = .019, d = .753. These results are summarized below.

The N250

The second component was analysed using the mean amplitude between 300 ms and 350 ms post stimuli onset. In this window, the repeated measures ANOVA revealed a main effect of scalp region, F(2.78,71.397) = 12.308, p < .001, η² = .436, and a voice condition x scalp region interaction, F(5.95,71.397) = 3.593, p = .004, η² = .028. Planned comparison with Bonferroni-corrected t-tests for the voice conditions within each given scalp region revealed a significant difference between TV and UV in MFC, t(12) = -3.479, p = .006 and LFC, t(12) = -3.506, p = .005, d = -.972 regions as well as between IFV and TV in LFC, t(12) = 3.12, p = .014, d = .865. No other difference was observed in this time window.

Since the main difference between TVs and UVs was the varying number of presentations in the course of the experiment, two ANOVAs with within-subject factors of voice condition and region were performed, respectively on the first and second halves of the experiment, to ascertain training effects. Only the ANOVA on the second half revealed significant interaction between voice condition and region F(4.428,53.137) = 3.544, p = .01, η² = .047. Again, Bonferroni-corrected planned comparison with voice conditions within each individual scalp region showed significant differences between TV and UV in MFC, t(12) = -3.215, p = .011, d = -.892 and LFC, t(12) = -2.624, p = .045, d = -.728 regions as well as between IFV and TV in LFC t(12) = 3.068, p = .016, d = .851.

The LPC

Finally, the LPC was investigated using a 50 ms sliding window of mean amplitudes between 450 ms and 850 ms post stimuli onset. Significant voice condition x scalp region interactions were found for 500–550 ms, F(6.719,80.632) = 3.411, p = .003, η² = .036, 600–650 ms, F(7.506,90.073) = 3.956, p < .001, η² = .043 and 650–700 ms, F(5.98,71.754) = 2.406, p = .036, η² = .041 time windows.

As before, planned comparison were used to compare the voice conditions within each individual scalp region using the Bonferroni correction for multiple comparisons. In the 500–550 ms time window, a significant difference between IFV and TV was found in RFC, t(12) = -2.854, p = .026, d = -.792 and LCP, t(12) = -3.188, p = .012, d = .884. For the 600–650 ms time window, there was a significant difference in mean amplitudes between IFV and both TV and UV in RFC, t(12) = -3.056, p = .016, d = -.847, t(12) = -3.418, p = .007, d = -.948, and in LCP t(12) = 3.221, p = .011, d = .893, t(12) = 4.192, p < .001, d = 1.163. A significant difference between IFV and UV was also found for this time window in MPC, t(12) = 2.920, p = .022, d = .810. Finally, the planned comparisons for the 650–700 ms window revealed significant differences between IFV and UV in MPC, t(12) = 2.782, p = .031, d = .7725 and LCP, t(12) = 2.608, p = .046, d = .723. As one can see by observing Fig 1, IFV was significantly different from both TV and UV in the early time window (200–250 ms) and for at least two windows within the LCP. Meanwhile, both TV and UV never differed significantly with the exception of the 300–350 ms time window during which UV was also different from IFV. The main regions involved were RFC for the P2 and the LCP, the MCF and LCF for the N250 and the MCP and LCP also for the LPC.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1.

Topographic representations of the ERP differences between (A) IFVs and TVs, (B) IFVs and UVs, and (C) TVs and UVs. Darkened areas and black dots represent regions and electrodes where voice conditions were significantly different. No significant difference were found on light-shaded topographies.

https://doi.org/10.1371/journal.pone.0250214.g001

To provide a broad picture of the results across the three time windows of interest, Fig 2 offers a summary of responses across the six regions. One can see that listeners’ neural responses to IFVs–where the speaker’s identity is known–stand out in the early time window in the RFC region on the P2, and in the late time window in the MPC, LCP and RFC on LPC. By contrast, listeners’ responses to TVs–for which the identity of the speaker is not known–appear in other regions, in this case at MFC and LFC regions in a mid-latency time window corresponding to the N250.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. ERPs on the six regions illustrating the effects of IFVs, TVs, and UVs in the three time windows of interest.

Distinct responses to IFV appear in an early window of 200–250 ms and were rightward as seen changing amplitudes at RFC, and also appeared in a late window of 500–650 ms where prolonged shifts appear in parietal sites at MPC and LCP and in frontal sites at RFC. For TVs and UVs, contrasting responses were found in a mid-late window of 300 to 350 ms, as seen in the differential responses at MFC and LFC.

https://doi.org/10.1371/journal.pone.0250214.g002