The study comprises two experiments distinguished by the signal processing strategy used for the manipulation of the modulation spectrum.

All listening subjects were young adults (college students) educated in the U.S. with normal hearing. Ten subjects participated in Experiment I; different set of subjects, 10 in number, participated in Experiment II. Although the number of listeners is smaller than typical for this type of study, their results (as described in Section “Results”) are consistent with each other.

The experimental corpus comprised 100 digit strings spoken fluently by a male speaker. Each string is a seven-digit sequence and is approximately 2 s long. It is uttered as a phone number in an American accent, i.e., a cluster of three digits followed by a cluster of four digits (for example: “two six two, seven O one eight”). It is a low perplexity corpus (a vocabulary of 11 words, 0–9, and O) but without contextual information. For each signal-manipulation condition, 80 stimuli (out of 100) were chosen at random and concatenated in a sequence: [alert tone] [digit string] [5-s long silence gap] [alert tone] …

Subjects performed the experiment in an isolated office environment (no other occupants) using headphones. There were two listening sessions for each signal-manipulation condition, Training and Testing. A training set contained 10 digit strings and a testing set contained 80 digit strings (approximately 12 min to complete). Training preceded testing; in the training phase, subjects had to perform above a prescribed threshold before proceeding to the testing phase. Subjects were instructed to listen to a digit string once and, during the 5-s long gap following the stimulus, to type into an electronic file the last four digits heard, in the order presented (always four digits, even those that she/he was uncertain about). The rational behind choosing the last four digits as target (as opposed to choosing the entire seven-digit string) was twofolded. First, it was an attempt to provide the opportunity for the presumed (cortical) theta oscillator to entrain to the input-rhythm prior to the occurrence of the target words (recall the inherent rhythm in the stimuli, being a seven-digit phone number uttered in an American accent). Second, it aimed at reducing the bias of memory load on the error patterns.

The human-subjects protocol for this study was approved by the Institutional Review Board of Boston University.

Data is presented in terms of error rate and normalized error rate. Error rate was calculated by using two distinct error metrics: (i) digit-error rate, defined as the number of digits erroneously registered divided by the total number of digits (i.e., 80 × 4 = 320 digits), in percent, and (ii) string-error rate, defined as the number of four-digit strings that – as a whole – were erroneously registered, divided by the total number of strings (i.e., 80 strings), in percent. Error rates were calculated per metric, per condition, per subject. The bar charts show the mean and the standard deviation across subjects. (The standard deviation here is the square root of the unbiased estimator of the variance.)

In order to exclude inter-subject variability, normalized error rates were calculated by normalizing the raw scores per metric, per subject, relative to the flat envelope condition (NoΘ in Experiment I; G = 1 in Experiment II). Normalized errors show the pattern of change in performance re flat envelope condition.

The conclusiveness of the experimental data was tested via analysis of variance (ANOVA; see Section “Statistical Analysis”).

Figure 8 shows the error rates (Figure 8A) and the normalized error rates (Figure 8B) for the Control, NoΘ, NoΘ + ChΘ, and NoΘ + GlbΘ conditions. As expected, the intelligibility of the Control stimuli is near perfect [see item 1 of Section “Experiment I: peak position coding (PPC) of critical-band envelopes”]. Comparing NoΘ to Control, errors jump from 0 to 7% (digit) and from 0 to 36% (string). For the task in hand (a low perplexity corpus and a limited memory load) such volume of errors is high. Re-inserting the ChΘ signal improves performance. Comparing NoΘ to NoΘ + ChΘ, errors drop from 7 to 4 (digit), and from 36 to 19 (string). Normalized NoΘ + ChΘ errors (re NoΘ) are 0.45 (digit) and 0.50 (string). Is the cause for this improvement the reactivation of the parsing path, the improved performance of the decoding path (due to extra spectro-temporal information still exists in ChΘ), or both?

Figure 9 provides a partial answer to this question. It shows errors, in percent, for the Control, NoΘ, and ChΘ (alone) conditions, with 0, 7, and 12 (digit) and 0, 36, and 50 (string), respectively. Intelligibility of ChΘ stimuli is worse than that of NoΘ stimuli, yet performance is better than chance, suggesting the existence of acoustic-phonetic information residue in the ChΘ stimuli, and indicating an imperfect disassociation between the roles of parsing and decoding.

In contrast, GlbΘ stimuli carry minimal (if any) acoustic-phonetic information. Comparing NoΘ to NoΘ + GlbΘ (Figure 8), errors drop from 7 to 5 (digit), and from 36 to 29 (digit); the corresponding normalized errors (re NoΘ) are 0.78 (digit) and 0.82 (string). As discussed in item 5 of Section “Experiment I: peak position coding (PPC) of critical-band envelopes,” this improvement is exclusively due to the recovered function of syllabic parsing.

Three observations are noteworthy. First, note the similarity in normalized error patterns between the digit- and string-error metrics. Second, the standard deviation bars in the normalized error patterns indicate strong degree of consistency across subjects for the NoΘ vs. NoΘ + ChΘ contrast. And third, a drop in the degree of consistency is noticed for the NoΘ vs. NoΘ + GlbΘ contrast. These observations will be quantified in the ANOVA presented in Section “Statistical Analysis.”

Figure 10 shows the error rates (Figure 10A) and the normalized error rates (Figure 10B) for the conditions G = 0, = 1, = 0.8, and = 0.5. As expected, the intelligibility of the G = 0 stimuli is near perfect. Comparing G = 1 to G = 0, errors jump from 0 to 4% (digit), and from 1 to 24% (string). For the task in hand (a low perplexity corpus and a limited memory load) such volume of error is high. Reducing the gain (i.e., reinstating critical-band envelope fluctuations) markedly improves intelligibility; comparing G = 1 to G = 0.8 (or G = 0.5), errors drop from 4 to 1 (or 0)% (digit), and from 24 to 7 (or 1)% (string). Normalized errors (re G = 1) are 0.07, 0.33, and 0.08 (digit), and 0.05, 0.33, and 0.06 (string), for G = 0, = 0.8, and = 0.5, respectively. As discussed in property 4 of Section “Experiment II: infinite-clipping (InfC) of critical-band signals,” this improvement is exclusively due to the recovered function of syllabic parsing. Note the consistency of normalized error patterns across subjects, and the similarity in normalized error patterns between the digit- and string-error metrics. This observation is quantified in the ANOVA presented next.

An ANOVA was used to quantify the statistical significance of the data illustrated in Figures 8–10. Three factors were used, system-type (PPC vs. InfC), condition ([Control/G = 0] vs. [NoΘ/G = 1] vs. [NoΘ + ChΘ/G = 0.5] vs. [NoΘ + GlbΘ/G = 0.8]) and error-metric (digit vs. string). Note that the variables in the condition factor, in particular [NoΘ + ChΘ/G = 0.5] and [NoΘ + GlbΘ/G = 0.8], were lumped somewhat arbitrarily.

Mauchly’s test for sphericity revealed that assumptions of sphericity were not violated. The three-way ANOVA revealed that there is a significant main effect of each of the factors: (i) system-type (F = 49.08, p < 0.0001), (ii) condition (F = 46.19, p < 0.0001), and (iii) error-metric (F = 118.19, p < 0.0001). It also revealed that there is a significant interaction between each of the pairs: (i) system-type and condition (F = 7.05, p = 0.0002), (ii) system-type and metric (F = 22.36, p < 0.0001), and (iii) condition and metric (F = 21.27, p < 0.0001). A post hoc Tukey/Kramer test, performed independently on each of the data sets and per error-metric, revealed the relationships summarized in Table 1.

The ANOVA reinforces the trends observed in Sections “Experiment I” and “Experiment II,” i.e., that the intelligibility of stimuli with flat critical-band envelopes is poor, and that the addition of extra information, restricted to the input syllabic rhythm alone, significantly improves intelligibility.

When considering the possible role of the temporal envelope of speech in speech perception, the term “envelope” is often refers to the envelope of the waveform itself, i.e., of the full-band signal (e.g., Ahissar et al., 2001; Giraud and Poeppel, 2012; Zion-Golumbic et al., 2012). In this study, the signal processing manipulations are performed on the critical-band outputs. The sole acoustic input available to the brain is, by necessity, the information conveyed by the auditory nerve. In the context of studying cortical mechanisms for speech perception, a more insightful approach would be to consider the information available to the brain at the cochlear output level. To exemplify the benefit of such approach consider the difference between the time-frequency representations caricatured in Figures 3B,C. Figure 3B depicts a minimalistic representation of the input rhythm at the cochlear output while Figure 3C depicts the input rhythm at the waveform level. If the hypothesized role of theta (i.e., tracking the input syllabic rhythm) is correct, does the neuronal mechanism that tracks the input rhythm (e.g., a neuronal PLL circuitry; Ahissar et al., 1997) exploits the information embedded in the curvature of the pulse contours of Figure 3B?

Chronologically, the PPC system was developed first. Conceptually, it aimed at improving the disassociation between the roles of parsing and decoding compared to the isolation provided by the strategy used by Chait et al. (2005) and Saoud et al. (2012). The InfC system emerged later, as a solution to shortcomings of the PPC system.

With the PPC system, Class-I stimuli are the NoΘ stimuli, composed of critical-band envelopes with null energy inside the 2- to 9-Hz modulation-frequency band. As such, these signals are stripped not only from input-rhythm information but also from a significant amount of acoustic-phonetic information (e.g., Houtgast and Steeneken, 1985; Drullman et al., 1994). One consequence of this deficiency is that the PPC strategy cannot be generalized to tasks with higher perplexity (see Section “Perplexity of Corpus” below). With the InfC system, on the other hand, the acoustic-phonetic information inside non-zero signal intervals for the G = 1 stimuli, as seen at the listener’s cochlear output, is the same as for the G = 0 corresponding stimuli [see property 3 of Section “Experiment II: infinite-clipping (InfC) of critical-band signals”], thus allowing a generalizing to tasks with higher perplexity.

With the PPC system Class-II stimuli, significantly more intelligible than the corresponding NoΘ stimuli, are the NoΘ + ChΘ stimuli. The existence of residual acoustic-phonetic information in the ChΘ signal [item 5 of Section “Experiment I: peak position coding (PPC) of critical-band envelopes”] implies that association between the functions of parsing and decoding still exists, thus preventing a decisive validation of the role of input rhythm in syllabic parsing per se. With the InfC system, on the other hand, the acoustic-phonetic information inside non-zero signal intervals, as seen in the listener’s cochlear output, remains the same for all values of G, e.g., for any prescribed degree of temporal envelope fluctuations [see property 3 of Section “Experiment II: infinite-clipping (InfC) of critical-band signals”], implying a clear disassociation between the roles of parsing and decoding.

One drawback of using semantically plausible material (such as TIMIT or the Harvard-IEEE sentences) in perceptual studies that measure word error rate is the ability of listeners to guess some of the words using contextual information. Semantically unpredictable sentences (SUS) make it more difficult for the listener to decode individual words on the basis of semantic context. For this reason, Ghitza and Greenberg (2009) used SUS material in which the sentences conformed to standard rules of English grammar but were composed of word combinations that are semantically anomalous (e.g., “Where does the cost feel the low night?” and “The vast trade dealt the task”). The SUS is a high perplexity corpus.

In preliminary trials with the PPC system an attempt was made to use the SUS corpus. Informal listening revealed that, because of the extensive removal of acoustic-phonetic information in generating the NoΘ (see Section “The PPC System vs. the InfC System”), NoΘ + ChΘ stimuli became hardly intelligible. This finding led to the selection of a seven-digit string corpus for this study – a low perplexity corpus but without context. On the other hand, informal listening to SUS stimuli generated by the InfC system showed that for a range of G values the stimuli are intelligible. Extending Experiment II to SUS material is beyond the scope of this study.

The behavioral data presented here show that intelligibility is improved by reinstating the input rhythm, either by adding a ChΘ (or a GlbΘ, to a lesser degree) signal to the NoΘ signal (Experiment I), or by reducing the Gain parameter from G = 1 to G = 0.5 or = 0.8 (Experiment II). What neuronal mechanism is capable of exploiting the re-inserted information in such an effective manner?

Our interpretation of the data argues for a crucial role of theta in syllabic parsing. We suggest that for stimuli with flat critical-band envelopes the tracking mechanism of the neuronal theta oscillator is deactivated. Consequently, the hierarchical window structure is situated in idle mode, i.e., not synchronized with the input, resulting in deterioration in performance. Reinstating the input-rhythm information (either by the PPC or by the InfC system) revives theta tracking hence the recovery of the synchronization between the window structure and the input, resulting in improved performance. This interpretation is consistent with the capability of the Tempo model to account for the behavioral data of Ghitza and Greenberg (2009), as detailed in Ghitza (2011).

Another possible interpretation of the data is in line with the classical view that assumes a direct role, in the decoding process per se, of neural onset responses to acoustic edges such as CV boundaries. According to this view phones are extracted from waveform segments (dyads) “centered” at markers triggered by acoustic edges. (It is worth recalling that a dyad is the acoustic reflection of the dynamic gesture of the articulators while moving from one phone to the next.) This view is supported by psychophysical reasoning, inferred from acoustics, on the importance of acoustically abrupt landmarks to speech perception (e.g., Stevens, 2002). Indeed, re-inserting the input rhythm with the PPC system produce signals with acoustic edges (NoΘ + ChΘ and NoΘ + GlbΘ; Figures 4F,G), which enable activation of neuronal circuitry sensitive to edges. Yet, two observations pose a challenge to the hypothesis that these edges are part of the decoding process: (i) Re-inserting GlbΘ signals, with pulses situated either at CV boundaries or at mid vowels, resulted in similar level of performance, and (ii) Reinstating temporal fluctuations by using the InfC system produce signals with a marginal presence of acoustic edges (see Figures 7D,E), yet with high intelligibility. These observations imply that the precise location of neural responses triggered by acoustic edges have little or no effect on intelligibility.

We suggest a different role to acoustic landmarks, in line with our interpretation of the data, in which they are part of the parsing (rather than decoding) process. We argue that neural responses triggered by acoustic landmarks serve as input to the mechanism that tracks the input syllabic rhythm. Note that this hypothesis includes all classes of acoustic landmarks (e.g., vocalic landmarks, glide landmarks, acoustically abrupt landmarks; Stevens, 2002). Given the prominence of vocalic speech segments in the presence of environmental noise, it may be that vocalic landmarks are more important than others in securing a reliable theta tracking. If so, a theta cycle is robustly aligned with [Vowel]–[Consonant-cluster]–[Vowel] syllables, and phones are decoded in temporal windows defined by the cycles of the beta, entrained to theta (Ghitza, 2011).