We examined 7 normal-hearing, right-handed monolingual native German speakers (3 female, 4 male). The age range was 22 to 27 years. None was on medication at that time. All participants signed an informed consent.

In this study, two types of stimuli were used to elicit electrophysiologic responses: speech stimuli and noise sound stimuli. The speech stimuli consisted of four monosyllabic words naturally produced by a male speaker, taken from Freiburger speech discrimination test [17]. The stimuli were Ei /a:i/, Bett /bεt/, Dieb /di:b/, and Pult /pult/, which mean egg, bed, thief and desk. The durations of the stimuli Ei, Bett, Dieb, and Pult were 544 ms, 430 ms, 501 ms, and 567 ms, respectively. The speech stimuli exhibited a bandwidth of 8 kHz and were digitalized with 14 bit at a sampling rate of 20,000 s^-1. Except Ei, all speech stimuli started with an initial stop consonant and were chosen with respect to their varying VOT. The VOTs were determined as VOT(Ei) = 0 ms, VOT(Bett) = 35 ms, VOT(Dieb) = 60 ms, and VOT(Pult) = 80 ms. The onset times of the closing consonants which succeed the vowels were determined as 215 ms with Bett, 390 ms with Dieb, and 395 ms with Pult.

The speech stimuli served as raw material for the synthesis of the noise stimuli. Noise stimuli were created by randomly multiplying the discrete samples of the speech stimuli with ± 1 and subsequent 8 kHz low pass filtering. Fig. (1) shows exemplarily the generation of the noise stimulus corresponding to Dieb. Thus, any phonetic information was removed from the speech stimuli, whereas the intensity was held constant. By this means, the synthesized noise stimuli Ei_N, Bett_N, Dieb_N, and Pult_N were generated.

Fig. (1). Speech stimulus Dieb and synthesized noise stimulus DiebN. Both stimuli exhibit the same temporal envelope but phonetic information is discarded in the noise stimulus.

Spectrograms of all presented stimuli are depicted in Fig. (2). The speech stimuli mainly consisted of frequencies lower than 2 kHz, whereas the noise stimuli were distributed over the entire bandwidth of 8 kHz.

Fig. (2). Spectral view of the presented stimuli. Top: Natural spoken monosyllabic words. Bottom: Synthetic noise sounds, derived from the monosyllabic words. Bandwidth of the stimuli is 8 kHz.

For equated presentation loudness, the applied stimuli were calibrated with respect to their root mean squares to a 1 kHz reference tone with an intensity level of 70 dB SPL (see [18]). Stimuli were presented in four blocks of 320 sounds each. Within the block, all stimuli were successively presented 40 times in the order of speech and corresponding noise stimulus (speech-noise pairs). Thus, each stimulus was presented 160 times. The interstimulus interval varied randomly between 1400 and 2100 ms. Subjects were tested in a soundproof booth. In order to minimize effects of the subjects’ state of attention on the applied stimuli, they were asked to watch a silent movie presented on a TV screen.

The EEG was derived with Ag/AgCl electrodes, which were integrated in a cap (Braincap, Brain Products GmbH, Gilching, Germany) with 30 fixed electrode positions (Fp1, Fp2, F3, F4, C3, C4, P3, P4, O1, O2, F7, F8, T3, T4, P7, P8, Fz, FCz, Cz, Pz, FC1, FC2, CP1, CP2, FC5, FC6, CP5, CP6, TP9, TP10). The electrode impedances were kept below 5 kOhm. For eye artefact rejection, an electro-oculogram was recorded by an electrode placed under the right eye. The EEG was recorded with Brain Amp-MR (Brain Products GmbH, Gilching, Germany) at a sampling rate of 500 Hz and digitalization of 16 bit.

During data acquisition, all channels were referenced to FCz. Offline, data were re-referenced to the mean of TP9 and TP10. The EEG was 0.13 Hz – 20 Hz band pass filtered with a slope of 12 dB/octave. The recording window included 100 ms prestimulus and 700 ms poststimulus time. Sweeps with artefacts measuring higher than 75 µV were rejected. The remaining sweeps were averaged separately for each stimulus and prestimulus baselined.

For each stimulus, grand mean waveforms were composed and grand mean latencies of the AEP components were visually identified. The individual peak latencies were determined in a time interval of the central latencies ± 24 ms. Amplitudes were calculated as a mean voltage at the 20 ms period centered at the individual peak latencies.

In the first step, statistical comparative tests of the AEP characteristics were carried out for both stimulus types, i.e. speech and noise stimuli, separately. In the second step, cortical responses to speech and corresponding noise sounds were analysed and compared among each other.

Preliminary one-way ANOVAs, which separately analysed N1 and P2 peak amplitudes, revealed significant amplitude differences between the electrode sites for both, speech and noise stimuli (p < 0.001, F_26,162 = 3.3 – 8.6). Newman-Keuls post hoc analysis revealed that with all stimuli, the greatest amplitudes were elicited in the central area of the scalp (i.e., channels Cz3, Cz, Cz4) for both, N1 and P2 amplitude. Hence, the following illustrations and statistics are restricted to channel Cz. Fig. (3a) shows the grand mean waveforms of the cortical responses to the four different speech stimuli. Morphologies of the speech-evoked potentials differed clearly across the stimuli. AEPs in response to Bett and Pult appear to reveal multiple overlapping response patterns. In all subjects, the predominant components were N1 and P2 in all responses. All speech stimuli elicited an N1 component which peaked around 113 ± 8 ms, and a P2 component that peaked in a range from 180 to 260 ms around 207 ± 36 ms.

Fig. (3). (a) Grand mean waveforms of speech stimuli elicited AEPs in response to Ei, Bett, Dieb, and Pult at channel Cz. (b) Grand mean waveforms of noise-sound elicited AEPs in response to EiN, BettN, DiebN, and PultN at channel Cz.

Fig. (3b) shows the grand mean waveforms of the cortical responses to the four different noise stimuli. The waveforms of the different noise sound elicited AEPs were similar. The predominant components in all responses were N1 and P2. All noise stimuli elicited an N1 component which peaked around 103 ± 8 ms, and a P2 component that peaked around 175 ± 11 ms.

In Tables 1 and 2, N1 and P2 latencies and amplitudes of both stimulus types are given. Table 3 shows F values of one-way ANOVAs, conducted to reveal significant differences of these characteristics. Newman-Keuls post hoc tests did not show consistent correlations within a subset of both stimulus types. As it is apparent from the higher levels of significance, speech evoked AEP characteristics diverged stronger among each other than those evoked by noise sounds.

For the analysis of hemispheric asymmetries, a paired t-test was conducted with N1-P2 interpeak amplitudes of opposite channel pairs T3 – T4, FC5 – FC6, and C3 – C4, continuing from lateral to central. The mean N1-P2 interpeak amplitude of T3 was significantly smaller than that of T4 for the speech stimuli but not for the noise stimuli, see t values in Table 4. While N1-P2 interpeak amplitudes at the positions FC5 and FC6 did not differ, an inversion from right to left hemispheric predominance occurred at the central positions C3, C4 for the speech stimuli (C3 > C4, see Table 4). Exemplarily, cortical responses to Dieb and Dieb_N across the entire scalp are depicted in Fig. (4), in order to illustrate the hemispheric asymmetries of the speech evoked potentials.

Fig. (4). Grand mean AEPs to Dieb (solid) and DiebN (dashed) over the entire scalp.

The AEP waveforms, in response to the speech sounds, were compared to those in response to the corresponding noise sounds. The time intervals, in which speech and corresponding noise responses significantly differed, are depicted in Fig. (5). Blocks of black colour mark temporal areas of significant differences (p<0.01) for each subject and each speech-noise pair. None of the speech-noise pair responses differs strongly during the initial 50 ms, as one can see from Figs. (5a-5d). The highest accumulations of areas of significant differences occur in the time window of 100 – 300 ms after stimulus onset.

Fig. (5). Temporal areas of significant differences between the responses to speech and noise stimuli at channel Cz for each subject (black areas). Level of significance is p<0.01.

Fig. (6) shows the grand mean AEP waveforms of the speech (dashed lines) and noise stimuli (dotted lines) as well as the difference signals (solid lines) speech – noise. In the difference signal, a negative component could be observed in a latency range of approximately 140 ± 15 ms. For each speech-noise pair, one-tailed t-tests (subjects x channels) showed that the occurring negative component was significantly smaller than zero. Latencies and amplitudes of this negative component are given in Table 5. Ei elicits the smallest latency, latencies in response to Bett and Dieb are similar and Pult elicits the longest latency. Fig. (7) shows the VOT of the four speech stimuli plotted against the peak latency of the negative component in the difference signal. A prominent correlation of VOT and peak latency is observable. Spearman rank correlation analysis revealed that the latencies were significant positively correlated with VOT (r = 0.66, p < 0.01).

Fig. (7). Box and whisker plot of VOTs of the diverse speech stimuli against the latencies of the negative components in the difference signal of the cortical responses. Dashed lines within the boxes display the median, the edges of the boxes display the quartiles. Whiskers indicate maximal and minimal latencies unless outliers ‘+’ occur (whiskers span 1.5 times the inter-quartile range).

Fig. (6). Grand mean waveforms at channel Cz of speech evoked responses (dashed), noise evoked responses (dotted) and the difference signal (solid).

In accordance with other studies (e.g. [6, 20, 21]), for both types of stimuli, the AEPs were maximal at central electrode sites. The different natural speech stimuli evoked distinct neural response patterns, partially showing multiple overlapping responses. These distinctions indicate an extended neural activity with the auditory processing of speech in contrast to noise sound processing. Considering previous studies, a plausible explanation are the variant VOTs of the presented speech stimuli which affect the N1 and P2 [8, 13, 22, 23].

Winkler et al. demonstrated that N1 latencies evoked by spectral-pitch and missing-fundamental tones do not differ [24]. Hence, the different pitches of the presented vowels are probably not the reason for the found variability in the speech evoked potentials, at least not regarding the N1. An influence of the cortical response to the closing consonant on the P2 is not expectable due to its relatively late onset time.

In contrast to noise elicited AEPs, hemispheric differences occurred with speech evoked AEPs in the present study. Hemispheric left-overbalanced asymmetries with speech perception are well known [25-27]. Left-overbalance is primarily reported in studies with attentive designs. However, focused auditory attention can selectively modulate automatic processing in auditory cortex and thus, affect AEPs [28].

Dehaene-Lambertz et al. demonstrated in a discrimination task that the left-hemispheric predominance with phoneme perception also can be achieved with appropriate non-speech stimuli [29]. Thus, specialization of left auditory cortex is not speech specific but depends on rapid spectrotemporal changes. However, in the present study, a right-hemispheric predominance was observable at temporal sites which inverted to a left-overbalance at central sites. This could be due to deviant locations in the activated auditory areas of both sides, resulting in different dipole orientations which underlie AEP derivations [30]. A more plausible explanation is a functional asymmetry of the auditory cortices. PET findings indicated complementary specializations of left and right auditory cortex’ belt areas [31]. Responses to temporal features were weighted towards the left, while responses to spectral features were weighted towards the right hemisphere.

Thus, the spectral changes of the consonant-vowel transitions may be the reason for the observed right hemispheric overbalance. In contrast, noise stimuli did not evoke asymmetric responses.

Although phonetic changes were present in the speech stimuli during the entire propagation of approximately 500 ms, AEPs of speech-noise pairs differed predominantly in the time window of the N1/P2 complex between 100 and 300 ms (Fig. 5), reflecting distinct onset processing.

Subtracting noise from speech elicited AEPs facilitated a direct comparison of the speech-noise pairs. Thus, the presented paradigm allows observing the neural processing of spectral-acoustic and phonetic information, since differences in stimulus duration and amplitude are eliminated. The comparison revealed a significant negative component in the difference signal that seems to correlate with the VOT of the presented stimulus (see Steinschneider et al. who found that synchronized activity in the auditory cortex is time locked with consonant release and voicing onset [32]). Since the speech evoked N1 and P2 appear somewhat later and heightened in contrast to the noise evoked, the present study suggests that the observed negative component denotes a second N1/P2 complex which is merged with the onset response. Therefore it is not directly visible. This second N1/P2 complex is suggested to reflect the onset of the vowel as an acoustic event.

This study demonstrated that rapid spectral changes in the perception of speech can be assessed electrophysiologically by means of a time-saving design. In case of modified cortical responses in subjects with central auditory processing disorders, this paradigm might serve as diagnostic tool for phonetic discrimination tasks since it reveals the neural processing of time-critical speech structures.