Stimulus and Task

Audio recordings of 2 detective stories by Sir Arthur Conan Doyle, “The yellow face” and “The Greek interpreter,” served as the to-be-attended and the to-be-ignored speech streams, respectively. Both stories were read by nonidentical male speakers. The speech streams were adjusted for root mean square intensity and cut into consecutive 30-s segments. In each 30-s segment, the speech streams were tagged with a sinusoidal amplitude modulation of 95 Hz (to-be-ignored stream) or 105 Hz (to-be-attended stream) modulation rate and a 50% modulation depth. Both speech streams were presented diotically, thus appearing to originate from a central sound source. Stimulus level was set to 70 dB(A).

The selective listening task is depicted in Figure ​Figure11A. Selective-listening trials, in which both speech streams interfered with each other, alternated with single-speaker trials, in which only the to-be-attended speech stream was present. The trial duration was set to 30 s for both conditions. During each trial, a target word was presented visually on a projection screen. The target word appeared 5 s after trial onset and was presented for 20 s. For the remaining time, a fixation cross was displayed to stabilize eye gaze.

Open in a separate window

Figure 1.

Selective listening task and behavior. (A) Participants performed a sustained selective listening task in which they attended to 1 of 2 simultaneously presented audio streams. Selective-listening trials alternated with single-speaker trials, in which only the to-be-attended speech stream was present. Listening performance was measured using a target word detection task. Target words occurring in the to-be-ignored speech stream (gray) had to be ignored. (B) The figure depicts the individual target detection accuracy (gray) as well as group averages (black; ±standard deviation) in single speaker and selective-listening trials of the experiment. (C) The duration of musical training and auditory working memory scores were associated with individual differences in the relative enhancement of target detection accuracy during selective listening.

Participants were instructed to press a button on a response pad whenever the target word occurred in the to-be-attended speech stream. Listeners were instructed not to respond for target words occurring in the to-be-ignored stream. The target word was changed after each task trial. In the selective-listening condition, the target word occurred within the to-be-attended and the to-be-ignored stream with equal probability. In the single-speaker condition, the target word was absent in half of the trials.

The task was split into 3 runs, each consisting of 10 selective-listening and 10 single-speaker trials. Runs started with a single-speaker trial to facilitate the initial identification of the to-be-attended speech stream. The to-be-attended stream and the to-be-detected target words were identical across participants. After finishing the experiment, participants were asked to freely recall the content of the to-be-attended story.

Additional Behavioral Testing

Prior to MEG recording, several other tests were carried out. A standard pure tone audiogram (frequency range: 0.125–8 kHz) was obtained from all participants. The pure tone average (PTA; i.e., the mean hearing threshold obtained at 500, 1000, 2000, and 4000 Hz, averaged over both ears) served as measure of individual hearing loss.

The Hearing in Noise test (HINT; Nilsson et al. 1994) was used to assess the 50% speech reception threshold for sentences in noise (SRT_Noise). We applied an adaptive 1-up-1-down staircase procedure in which the presentation level of the speech stimuli was varied adaptively in steps of 3, 1, or 0.5 dB. The step size was decreased after the second and fourth turning point. The noise level was set to 65 dB(A). The signal-to-noise ratio (SNR) in the first trial was set to 6 dB. The staircase consisted of at least 23 trials and 10 turns. A training run was used to familiarize subjects with the task. The average SNR obtained across all trials following the fourth turning point served as measure of SRT_Noise.

Complementing the HINT measures, the speech reception threshold for sentences in a concurrent speech background (SRT_Speech) was obtained using a task design adapted from Swaminathan et al. (2015). In each trial, participants listened to 1 of 3 simultaneously presented 5-word sentences, uttered by 3 nonidentical male speakers. All sentences were taken from a closed set matrix corpus developed by Kidd et al. (2008) and had an identical grammatical structure (i.e., <name><verb><number><color><noun>). In each trial, 3 speakers were chosen randomly from a pool of 11 male voices. Participants were instructed to repeat the sentence beginning with the word “Mike.” The target sentence was presented at the fixed level of 55 dB(A). SNR in the first trial was set to 15 dB. The level of the competing sentences was adjusted parametrically in steps of 4 or 1 dB (after 2 turns) in a 1-up-1-down staircase procedure. The staircase consisted of at least 23 trials and 10 turns. SRT_Speech was calculated from the average SNR of all trials following the fourth turning point. A training run was used to familiarize participants with the task.

Performance in a tonal sequence manipulation task (Foster et al. 2013; Albouy et al. 2017) was used as an index of auditory working memory performance. In each trial of this task, a 3-tone piano sequence was presented. After a 2-s retention period, the sequence was repeated in reversed temporal order. Participants had to judge whether a single tone was changed in the second presentation (in positions 1–3, pitch change of 2 or 3 semi tones equally distributed). The pitch change preserved the melodic contour of the sequence. The task consisted of 108 trials, half of them containing a pitch change. The presentation order was controlled in such a way that similar trial types (same, different) could not be presented more than 3 times in a row. The auditory working memory score was computed as the percentage of correct judgements given across all trials.

Behavioral Data Analysis

Target word detection accuracy in single-speaker and selective-listening trials of the selective listening task was quantified using the F1 score (Van Rijsbergen 1979). The F1 score represents the harmonic mean of recall (i.e., number of detected target words divided by the total number of targets words) and precision (number of detected target words divided by the total number of button presses). Group-level differences in target word detection accuracy between single-speaker and selective-listening conditions were assessed using a paired t-test.

To account for variability in baseline target detection performance between participants (Fig. ​(Fig.11B), the difference in F1 scores obtained in the single-speaker and selective-listening trials (F1) served as measure of individual selective listening performance in our study. We tested for linear relationships between F1 and audiological measures, the duration of musical training, and auditory working memory scores using robust regression with default settings as implemented in MATLAB (i.e., bisquare weighting, tuning constant set to 4.685). Robust regression was preferred to ordinary least-squares regression to reduce the effect of outliers on the parameter estimation. The goodness of the robust regression fit is stated in terms of R² values. Regression slopes are considered to differ significantly from zero when passing a statistical threshold of P_FDR < 0.05, corrected for multiple comparisons using the Benjamini-Hochberg method for false discovery rate correction (5 tests included; Benjamini and Hochberg 1995).

Robust regression was used to assess potential relationships between musical training and both auditory working memory scores and audiological measures. A significance criterion of P_FDR < 0.05 was applied (FDR correction over 4 tests). Given that previous studies suggest an increase of auditory working memory performance with increasing levels of musical training, we tested for positive relationships only. A Sobel test (Sobel 1982; MacKinnon et al. 1995) was performed to test whether auditory working memory acts as a mediator to convey effects of musical training on selective listening performance.

MEG Data Acquisition

MEG data were acquired during the selective listening task using a 275-channel whole-head MEG system (CTF/VMS, Port Coquitlam, British Columbia, Canada). Data were recorded with a sampling rate of 2400 Hz, an antialiasing filter with a 600 Hz cut-off, and third-order spatial gradient noise cancellation. Horizontal and vertical electrooculograms, and an electrocardiogram were acquired with bipolar montages. The head position inside the MEG sensor helmet was determined with coils fixated at the nasion and the preauricular points (fiducial points). For anatomical registration with anatomical MRI data, the spatial positions of the fiducial coils and of about 150 scalp points were obtained using a 3D digitizer system (Polhemus Isotrack, Polhemus, Colchester, VT, USA). Participants were seated in upright position in a sound-attenuated, magnetically shielded recording room. Auditory stimulation was delivered via insert earphones (E-A-RTONE 3A, 3M, St. Paul, MN, USA). The earphones were equipped with customized prolonged air tubes (1.5 m) to increase the distance between audio transducers and MEG sensors, thus minimizing potential stimulation artifacts in the MEG signal, which was confirmed with pilot testing using a foam head.

MEG Data Preprocessing

MEG data preprocessing was performed in Brainstorm (Tadel et al. 2011), in observance of good-practice guidelines (Gross et al. 2013). Line noise artifacts were removed using notch filters at the power frequency (60 Hz) and its first 3 harmonics. Artifacts related to eye movements and cardiac activity were pruned from the data using independent components analysis. For this procedure, a copy of the data was offline filtered between 1 and 40 Hz. A principal component analysis was performed to reduce the dimensionality of the MEG data to 40 dimensions and 40 independent components were computed using the extended infomax algorithm implemented in Brainstorm. The demixing matrix obtained from this procedure was applied to the original unfiltered MEG dataset and independent components which sensor topography and/or time series reflected eye blinks, lateral eye movements, and cardiac activity were removed. No further data cleaning was performed.

The participant’s structural T1 MRI images were automatically segmented and labeled using Freesurfer (Dale et al. 1999; Fischl et al. 1999, 2004), and coregistered to the MEG data in Brainstorm using the digitized head points. An OPENMEEG boundary element method head model and a surface-based minimum norm source model with depth weighting (order: 0.5; maximal amount: 10) was computed for each participant, also using Brainstorm. Dipole orientation was constrained to be normal to the cortical surface. Noise covariance matrices for the source reconstruction process were estimated from 2-min empty room recordings obtained for each participant.

FFR Analysis

The ongoing speech signal was tagged with an amplitude modulation. Previous work shows that the ongoing auditory response locks to such periodic signal amplitude fluctuations, resulting in spectral peaks at the modulation frequency (Picton et al. 2003; Bharadwaj et al. 2014). Individual differences in spectral power at the amplitude modulation rate have been related to behavioral measures of auditory temporal acuity (Purcell et al. 2004; Bharadwaj et al. 2015), suggesting that the FFR power can serve as a neural marker of auditory temporal coding fidelity. Please note that this subtype of the FFR is also known as envelope following response (Shinn-Cunningham et al. 2017).

MEG data demonstrate that FFRs can be measured not only at brainstem and midbrain levels, but also from auditory cortex (Schoonhoven et al. 2003; Coffey et al. 2016). We here obtained cortical FFRs from left and right Heschl’s gyrus. A copy of the cleaned MEG sensor data was bandpass-filtered between 60 and 120 Hz, downsampled to 600 Hz sampling rate, and epoched into continuous 2-s intervals. Each of the resulting epochs contained 190 or 210 full cycles (phase onset: 0) of the sinusoidal 95 and 105 Hz amplitude modulations applied to the speech stimuli. Averaging across epochs and conditions is thought to preserve the entrained oscillation at the modulation frequency while attenuating stimulus-independent intrinsic oscillations. MEG sources of the condition averages (single-speaker/selective-listening trials) were projected into the standard MNI space using Brainstorm. Welch’s power spectral density was estimated for all elementary sources within left and right Heschl’s gyrus regions and subsequently averaged within hemispheres. Anatomical labels were obtained automatically from Freesurfer, based on the Desikan-Killiany atlas (Desikan et al. 2006). To account for overall power differences across individuals, the spectral power at the tagged modulation frequencies is reported relative to the individual baseline power (i.e., mean spectral power between 60 and 120 Hz).

A 2 × 3 repeated-measures ANOVA tested for differences in FFR power related to hemisphere (2 levels: left/right) and listening condition (3 levels: single/to-be-attended/to-be-ignored stream). Post hoc paired t-tests were used to investigate the statistically significant main effect of listening condition in more detail. Results of the post hoc t-tests were reported as statistically significant when passing a threshold of P_FDR < 0.05 (FDR correction over 3 tests). Pearson correlation was used to investigate the stability of individual differences in FFR power across hemispheres (P_FDR < 0.05, FDR correction over 3 tests).

Since previous work suggests that selective attention may modulate FFR amplitudes (Lehmann and Schönwiesner 2014), FFR power measured during the single-speaker phase, which lacks any attention manipulation, served as unbiased measure of auditory coding fidelity. FFR power was averaged over both hemispheres for further analyses. Linear relationships between individual FFR power and the duration of musical training or task performance were analyzed using robust regression. Based on our hypotheses, we only tested for positive relationships. Results are reported as statistically significant for P_FDR < 0.05 (FDR correction over 2 tests).

Cortical FFR Analysis

Figure ​Figure22A depicts the power spectrum density estimates obtained from left and right Heschl’s gyrus sources in the single-speaker condition and during selective listening. The power spectra show bilateral peaks at the amplitude of the modulation frequencies applied to the to-be-attended (105 Hz) and to-be-ignored (95 Hz) speech streams. Figure ​Figure22B illustrates the mean power spectral density at 105 Hz for each MEG sensor in the single-speaker condition. The power topography is consistent with the pattern of MEG activity generated by an auditory cortex source (Coffey et al. 2016), providing evidence that the measured FFR response is primarily generated in cortical auditory sensory areas.

Open in a separate window

Figure 2.

FFR analysis: The single/to-be-attended (Sin/Att) and the to-be-ignored (Ign) speech stream were tagged with an amplitude modulation of 105 or 95 Hz modulation frequency. (A) The spectral power induced at the modulation frequencies was investigated in surface ROIs covering left and right Heschl’s gyrus. The figure depicts the individual power spectrum for each participant, with response peaks emerging at both modulation frequencies. (B) The group-level sensor topography of the power spectral density, obtained in the single-speaker condition at 105 Hz (i.e., the modulation frequency of the speech stream), is consistent with the pattern of MEG activity typically generated by an auditory cortex source. (C) FFR power varied significantly as between hemispheres (left/right Heschl’s gyrus) and as a function of listening condition (Sin = single-stream at 105 Hz, Att = to-be-attended stream at 105 Hz, Ign = to-be-ignored stream at 95 Hz). The figure shows individual data (gray) and group averages (black; ± standard deviation). Significant differences between conditions are marked by asterisks. (D) The FFR power obtained in the single stream condition, averaged over both hemispheres, served as a measure of temporal coding fidelity in auditory cortex. This measure was significantly related to the duration of musical training, but did not significantly explain individual differences in task performance.

A repeated-measures ANOVA was used to test for effects of hemisphere (left/right Heschl’s gyrus) and condition (single stream at 105 Hz/to-be-attended stream at 105 Hz/to-be-ignored stream at 95 Hz) on FFR power. The analysis revealed a main effect of listening condition (F[2,38] = 5.14, P = 0.011) as well as a significant main effect of hemisphere (F[1,19] = 5.32, P = 0.032), but no condition-by-hemisphere interaction (F[2,38] = 1.63, P = 0.210). As shown in Figure ​Figure22C, FFR power was reduced for the to-be-attended stream (4.6 ± 4.5 dB) as compared with both the single speaker (6.1 ± 5.4 dB; P_FDR = 0.028) and to-be-ignored speech stream (5.6 ± 5.0 dB; P_FDR = 0.048). No differences in FFR power were observed between single-speaker and to-be-ignored streams (P_FDR = 0.426). In line with data by Coffey et al. (2016), FFR power was overall higher in right (5.8 ± 5.1 dB) than in left auditory cortex (4.6 ± 4.5 dB). In all listening conditions, individual FFR responses were however highly correlated across hemispheres (single stream: r = 0.98, P_FDR < 0.001; to-be-attended stream: r = 0.65, P_FDR = 0.002; to-be-ignored stream: r = 0.73, P_FDR < 0.001).

The FFR power at 105 Hz as obtained in the single-stream condition, averaged over both hemispheres, served as a measure of temporal coding fidelity in auditory cortex. A robust regression analysis revealed that the individual FFR power was positively related to the duration of musical training (R² = 0.21, slope = 0.40 dB/year, P_FDR = 0.041, one-tailed; see Fig. ​Fig.22D), as predicted by previous studies. In contrast to musical training and working memory scores, FFR power was not significantly related to F1 scores in the selective listening task (R² = 0.11, P_FDR = 0.074, one-tailed). In an exploratory analysis, we subsequently confirmed that a similar relationship between FFR power and musical training can qualitatively also be observed for the to-be-attended speech stream (R² = 0.15, slope = 0.28 dB/year, P = 0.047, uncorrected, one-tailed). No such effect was found for the to-be-ignored speech stream (R² = 0.08, P = 0.109, uncorrected, one-tailed).

Musical Training is Associated With Better Selective Listening Performance

Overall, our participants performed worse in the presence of competing speech, as expected. Yet, the deleterious effect of the to-be-ignored speech stream on task performance decreased with increasing duration of musical training (Fig. ​(Fig.11C). This result suggests a higher resilience against auditory distractors in musically trained individuals during cocktail party listening. While multiple studies reported a musician advantage when listening to speech in noise (for a review, see Coffey, Mogilever et al. 2017), previous work on selective listening in cocktail party settings provided no consistent evidence for beneficial effects of musical training (Boebinger et al. 2015; Swaminathan et al. 2015; Baskent and Gaudrain 2016; Clayton et al. 2016). Most of these studies however tested for musical training-related effects on speech reception thresholds in speech backgrounds, whereas our experiment assessed performance during sustained selective listening at suprathreshold level. It is noteworthy that when applying the experimental design introduced by Swaminathan et al. (2015) and no spatial separation of sound sources, we observed no effect of musical training on speech reception thresholds in a speech background. This observation suggests that our present findings of a beneficial effect of musical training in cocktail party listening cannot be explained by changes in speech reception thresholds in speech backgrounds.

Similarly, our data provide no additional evidence for reduced speech-in-noise perception thresholds in musicians. Unlike the majority of studies reporting such effects, we however did not compare groups of trained musicians and nonmusicians (Coffey, Mogilever et al. 2017) but treated musical training as a continuous covariate in a heterogenous group of participants, including both musicians and nonmusicians.

Effect of Musical Training on Cortical Speech Tracking

During selective listening, speech envelope reconstruction of the to-be-ignored speech from MEG source time courses was degraded with respect to the attended speech stream (Fig. ​(Fig.33B). This result is in accord with previous work applying similar reconstruction approaches, and in line with the view that the attended speech stream is more strongly represented in the ongoing neural response than ignored speech (Ding and Simon 2012a, 2012b; Mesgarani and Chang 2012; Puschmann et al. 2017). Recent work suggests that attention-related modulations on cortical representations of speech occur at the level of auditory cortex, but only after initial sensory processing of the acoustic scene (Puvvada and Simon 2017).

In our study, robust cortical tracking of the to-be-attended speech stream was observed in auditory sensory regions, motor and somatosensory cortex, and inferior frontal brain regions (Fig. ​(Fig.33E). In contrast, tracking of the to-be-ignored speech stream was—on the group level—restricted to auditory cortex, in spite of possible field spread of MEG source imaging (Brodbeck et al. 2018). These observations are in line with electrocorticographic recordings reported by Zion Golumbic et al. (2013).

Previous work on cocktail party listening provide evidence for a link between selective listening performance and the cortical tracking of attended speech (O’Sullivan et al. 2015; Puschmann et al. 2017). The intelligibility of speech in noise has also been related to the accuracy of cortical speech envelope tracking (Ding and Simon 2013). Further, behavioral benefits of musicians in speech in noise perception were reported to be associated with a superior encoding of stimulus information, yielding more robust speech representations at both brainstem and cortical levels, including auditory sensory as well as premotor and frontal areas (Parbery-Clark, Skoe and Kraus 2009; Du and Zatorre 2017). Based on these findings, we expected to observe a positive relationship between the duration of musical training and the tracking of the to-be-attended speech stream. However, although our data provide some evidence for the previously reported relationship between the tracking of attended speech and behavior, effects of musical training on the envelope reconstruction accuracies obtained for the to-be-attended stream were relatively weak in our dataset. Instead, we observed a strong positive relationship between the individual duration of musical training and the tracking of the to-be-ignored speech stream (Fig. ​(Fig.33C). Further, a longer duration of intense musical training was associated with a more balanced cortical representation of both speech streams (Fig. ​(Fig.33D).

While positive effects of musical training on the tracking of the ignored speech were most robust over the left auditory cortex, we observed similar tendencies in motor and somatosensory cortices, the inferior parietal lobe, and inferior frontal brain regions (Fig. ​(Fig.44B). Increased tracking of ignored speech was persistent up until late processing time windows, suggesting that the increased resilience against distractor interference in musically trained subjects did not rely on an efficient filtering of the attended information (Fritz et al. 2007). Our findings are consistent with and extend recent fMRI results by Du and Zatorre (2017) on speech-in-noise processing in musicians and nonmusicians. Their data showed that musicians exhibit a more robust multivariate encoding of speech features in the auditory cortex and adjacent temporal regions in high signal-to-noise conditions, but also in motor and somatosensory cortices, and in parts of the inferior frontal lobe, especially as signal-to-noise conditions become more difficult. We here demonstrate that, in musically trained individuals, the same brain regions also reveal more robust temporal envelope representations of the to-be-ignored speech during cocktail party listening.

One possible explanation for the ability of musically trained persons to uphold representations of both attended and ignored sound streams is that musical training and performance typically require not only segregating one sound from a background, but also attending to and integrating multiple different musical streams (Disbergen et al. 2018). However, it is unknown how increased tracking of the to-be-ignored speech stream may benefit listening success on the attended channel. On the one hand, keeping track of both the attentional foreground and background may facilitate predictions of how both streams evolve over time and, thus, stabilize auditory stream segregation during cocktail party listening (Elhilali and Shamma 2008). In consequence, intrusions of the irrelevant speaker may be limited, thus reducing the overall listening effort. On the other hand, it may enable predictions on the temporal progression of the to-be-ignored stream, allowing better anticipation of time intervals of low acoustic energy in the noise background. However, although it is widely believed that listeners tend to “listen in the dips” of the noise background to improve speech perception, previous work does not provide strong evidence for effects of masker predictability on listening success (Cooke 2006; Jones and Litovsky 2008).

Relationship to Low-Level Sensory Processing and Auditory Working Memory

Musical training was previously demonstrated to benefit both low-level sensory processing of auditory information and higher cognitive functions, in particular auditory working memory (Kraus et al. 2012; Strait and Kraus 2014). In line with these findings, we also observed positive relationships between the duration of musical training and both auditory working memory performance and cortical FFR, which can be seen as a marker of temporal coding fidelity of low-level auditory processing. Differences in temporal coding fidelity and in auditory working memory however affected the cortical tracking of speech in different ways.

Individual differences in FFR power were positively related to the cortical tracking of the to-be-ignored stream. In contrast to effects found for musical training and working memory scores, there was no significant change in the relative cortical representation of attended and ignored speech with increasing FFR power. This suggests that musical training-related changes in FFR power improve the cortical tracking of both streams similarly. The association between FFR power and speech envelope tracking was strongest at early time points (i.e., the 15–80 ms time window) and in left auditory sensory brain regions, indicating that superior temporal coding of auditory information enabled more robust tracking at early stages of auditory sensory processing (Fig. ​(Fig.44B). This effect may be related to a higher fidelity of stimulus representations within the auditory system per se (Strait et al. 2014), but also to a superior stream segregation, resulting in more stable representations of both speech streams. In our task, participants presumably relied on spectral cues for early input-based stream segregation. Previous work provides evidence for a link between the enhanced temporal coding of auditory information in musicians and fine pitch discrimination abilities (Bidelman et al. 2011; Coffey et al. 2016; Coffey, Chepesiuk et al. 2017). Also, musicians were reported to show superior pitch-based concurrent stream segregation (Zendel and Alain 2008).

Similar to the duration of musical training, auditory working memory was a good predictor of behavioral performance (Fig. ​(Fig.11C). On the neural level, auditory working memory performance was positively related to the cortical tracking of the to-be-ignored, but not of the to-be-attended, speech stream, similar to the effect of musical training. Working memory-related modulations of speech tracking were strongest at intermediate time lags (i.e., the 90–175 ms time window), and primarily detected in left superior parietal and anterior cingulate regions, with similar trends in the left motor cortex and left inferior frontal lobe (Fig. ​(Fig.44A,B). This observation suggests that auditory working memory, in contrast to the fidelity of temporal information coding, does not amplify the early cortical representation of speech in auditory sensory areas, but rather strengthens the representation of the to-be-ignored speech input at later processing stages.

Auditory working memory was assessed using a tonal sequence manipulation task (Foster et al. 2013; Albouy et al. 2017), measuring a nonverbal tonal component of auditory working memory (Schulze and Koelsch 2012). Still, the tonal auditory working memory measure was associated with both the individual outcome of the verbal behavioral task and cortical speech tracking. Other studies however provided evidence for a similar musical training-related increase in verbal auditory working memory (Parbery-Clark, Skoe and Kraus 2009; Strait et al. 2012), suggesting that musical training can benefit different components of auditory working memory. Tonal and verbal working memory networks were previously reported to share a set of common core structures (Schulze et al. 2011), but can also be selectively impaired (Albouy et al., 2013).

Recent MEG data demonstrated that the motor system is involved in generating temporal predictions of auditory inputs to facilitate auditory stream segregation (Morillon and Baillet 2017). In the context of speech processing, delta-band inputs from left inferior frontal cortex and left motor cortex into left auditory cortex were found to be associated with increased neural coupling with the speech envelope in auditory cortex (Park et al. 2015). Similarly, cortical oscillations in orbitofrontal regions were shown to modulate the entrainment of auditory cortex by the speech envelope (Keitel et al. 2017). Taken together, these findings provide evidence for top-down predictive signaling from frontal and motor cortices directed to the auditory cortex, leading to superior speech envelope tracking. The superior parietal regions were, in contrast, previously reported to receive bottom-up signals from auditory sensory areas, related to the neural tracking of the speech envelope (Keitel et al. 2017).

Our data show that subjects with superior auditory working memory performance maintain increased representations of the ignored speech stream during selective listening, and that this is associated with superior parietal lobe areas, the left motor cortex, and left inferior frontal regions. We suggest that the latter brain regions generate input predictions of both the attended and ignored speech streams during cocktail party listening, thus facilitating ongoing stream segregation and, potentially, listening “in the dips” of the perceptual background. Listeners with inferior auditory working memory performance may not have the capacity to keep and process representations of both input streams, and therefore may only rely on predictions of the to-be-attended stream. However, future work is necessary to test this hypothesis specifically and in greater detail. Also, it remains open whether prediction signals of attended and ignored speech input are generated independently or interdependently of each other.