Structured Neuronal Encoding and Decoding of Human Speech Features

ARTICLE

Received 13 Feb 2012 | Accepted 6 Jul 2012 | Published 21 Aug 2012 DOI: 10.1038/ncomms1995 Structured neuronal encoding and decoding of human speech features

Ariel Tankus1,2, Itzhak Fried1,3,4 & Shy Shoham2

Human speech sounds are produced through a coordinated movement of structures along the vocal tract. Here we show highly structured neuronal encoding of vowel articulation. In medial– frontal neurons, we observe highly specific tuning to individual vowels, whereas superior temporal gyrus neurons have nonspecific, sinusoidally modulated tuning (analogous to motor cortical directional tuning). At the neuronal population level, a decoding analysis reveals that the underlying structure of vowel encoding reflects the anatomical basis of articulatory movements. This structured encoding enables accurate decoding of volitional speech segments and could be applied in the development of brain–machine interfaces for restoring speech in paralysed individuals.

1 Department of Neurosurgery, David Geffen School of Medicine, and Semel Institute for Neuroscience and Human Behavior, University of California, Los Angeles, California 90095, USA. 2 Faculty of Biomedical Engineering, Technion-Israel Institute of Technology, Technion City, Haifa 32000, Israel. 3 Functional Neurosurgery Unit, Tel-Aviv Medical Center, Tel-Aviv 64239, Israel. 4 Sackler Faculty of Medicine, Tel-Aviv University, Tel-Aviv 69978, Israel. Correspondence and requests for materials should be addressed to I.F. (email: [email protected]) or to S.S. (email: [email protected]). nature communications | 3:1015 | DOI: 10.1038/ncomms1995 | www.nature.com/naturecommunications © 2012 Macmillan Publishers Limited. All rights reserved. ARTICLE nature communications | DOI: 10.1038/ncomms1995

he articulatory features that distinguish different vowel sounds to distinguish between auditory- and speech-related neuronal are conventionally described along a two-dimensional coor- activations, we analysed only the 606 units that did not respond to Tdinate system that naturally represents the position of the auditory stimuli. A unit was considered speech-related if its firing highest point of the tongue during articulation, for example, in the rate during speech differed significantly from the pre-cue baseline International Phonetic Alphabet (IPA) chart1. The two axes of this period (see Methods). Overall, 8% of the analysed units (49) were system are height (tongue vertical position relative to the roof of the speech-related, of which, more than a half (25) were vowel-tuned, mouth or the aperture of the jaw) and backness (tongue position showing significantly different activation for the five vowels (see relative to the back of the mouth). How is the structured articu- Supplementary Fig. S1). latory production encoded and controlled in brain circuitry? The gross functional neuroanatomy of speech production was described Sharp and broad vowel tuning. Two areas commonly activated by multiple imaging, lesion and stimulation studies2,3, and includes in speech studies2, the STG and a medial–frontal region overlying primary, supplementary and pre-motor areas, Broca’s area, superior the rostral anterior cingulate and the adjacent medial orbitofrontal temporal gyrus (STG), anterior cingulate cortex (ACC) and other cortex (rAC/MOF; Brodmann areas 11 and 12; See Supplementary medial–frontal regions2–4. The temporal dynamics ofcollective neu- Fig. S2 for anatomical locations of the electrodes), had the high- ral activity was studied in Broca’s area using local field potentials2,5. est proportions of speech-related (75% and 11%, respectively) and However, the basic encoding of speech features in the firing patterns vowel-tuned units (58% and 77% of these units). In imaging and of neuronal populations remains unknown. electrocorticography studies, the rACC was shown to participate in Here, we study the neuronal encoding of vowel articulation speech control2,6, the orbitofrontal cortex in speech comprehension in the human cerebrum at both the single-unit and the neuronal and reading7, and the STG in speech production at the phoneme population levels. At the single-neuron level, we find signatures level8. Involvement of STG neurons in speech production was also of two structured coding strategies: highly specific, sharp tuning observed in earlier single-unit recordings in humans9. We analysed to individual vowels (in medial–frontal neurons) and nonspe- neuronal tuning in these two areas and found that it had divergent cific, sinusoidally modulated tuning (in the STG). At the neural characteristics: broadly tuned units that responded to all vowels, population level, we find that the encoding of vowels reflects the with a gradual modulation in the firing rate between vowels, com- underlying articulatory movement structure. These findings may prised 93% of tuned units in STG (13/14) but were not found in have important implications for the development of high-accuracy rAC/MOF (0/10), whereas sharply tuned units that had significant brain–machine interfaces for the restoration of speech in paralysed activation exclusively for one or two vowels comprised 100% of the individuals. tuned rAC/MOF units (10/10) but were rare in STG (1/14). Figure 1 displays responses of five sharply tuned units in rAC/ Results MOF, each exhibiting strong, robust increases in their firing rate Speech-related neurons. Neuronal responses in human temporal specifically for one or two vowel sounds, whereas for the other vow- and frontal lobes were recorded from 11 patients with intractable els firing remains at the baseline rate. For example, a single unit in epilepsy monitored with intracranial depth electrodes to identify the right rostral anterior cingulate cortex (rACC) (Fig. 1, top row) seizure foci for potential surgical treatment (see Methods). elevated its firing rate to an average of 97 spikes/s when the patient Following an auditory cue, subjects uttered one of five vowels (a/a/, said ‘a’, compared with 6 spikes/s for ‘i’, ‘e’, ‘o’ and ‘u’ (P < 10 − 13, one- e/ε/, i/i/, o/o/ and u/u/) or simple syllables containing these vowels sided two-sample t-test). Anecdotally, in the first two trials of this (consonant + vowel: da/da/, de/dε/, di/di/, do/do/ and du/du/…). example (red arrow) the firing rate remained at the baseline level, We recorded the activity of 716 temporal and frontal lobe units. As unlike the rest of the ‘a’ trials; in these two trials, the patient wrongly this study focuses on speech and owing to the inherent difficulty said ‘ah’ rather than ‘a’ (confirmed by the sound recordings).

a e i u o es/s 20

spik 1 s

Figure 1 | Sharply tuned medial–frontal (rAC/MOF) units. Raster plots and peri-stimulus time histograms of five units during the utterance of the five vowels a, e, i, u and o. For each of the units, significant change in firing rate from the baseline occurred for one or two vowels only (Methods). Red vertical dashed lines indicate speech onset. All vertical scale bars correspond to firing rates of 20 spikes/s.

a a e i u o b 30

25 es/s]

ing rate [spik 40 erage fir

Av 5 15 es/s 10 0

spik o a e i u o 1 s

Figure 2 | Broadly tuned STG units. (a) Raster plots and peri-stimulus time histograms during the utterance of the five vowels a, e, i, u and o.S ignificant change in firing rate from the baseline occurred for all or most vowels, with modulated firing rateM ( ethods). Red vertical dashed lines indicate speech onset; vertical bars, 10 spikes/s. (b) Tuning curves of the respective units in a over the vowel space, showing orderly variation in the firing rate of STG units with the articulated vowel.

A completely different encoding of vowels was found in the STG, n = 6 cross-validation runs; Supplementary Table S1), and up to where the vast majority of tuned units exhibited broad variation of 100% when decoding pairs of vowels (Fig. 3a). Next, we selected the response over the vowel space, during the articulation of both the vowel ordering that leads to band-diagonal confusion matrices vowels (Fig. 2a) and simple syllables containing these vowels (Sup- (Fig. 3b). Interestingly, this ordering is consistent across different plementary Fig. S3a). This structured variation is well approximated neuronal subpopulations (Fig. 3b and Supplementary Fig. S4) and by sinusoidal tuning curves (Fig. 2b and Supplementary Fig. S3b) exactly matches the organization of vowels according to their place analogous to the directional tuning curves commonly observed and manner of articulation as reflected by the IPA chart (Fig. 3c). As in motor cortical neurons10. Units shown in Fig. 2 had maximal the vowel chart represents the position of the highest point of the responses (‘preferred vowel’, in analogy to ‘preferred direction’) tongue during articulation, the natural organization of speech fea- to the vowels ‘i’ and ‘u’, which correspond to a closed articulation tures by neuronal encoding reflects a functional spatial-anatomical where the tongue is maximally raised, and minimal (‘anti-preferred’) axis in the mouth. response to ‘a’ and ‘o’ where it is lowered. Discussion Population-level decoding and structure. Unlike directional tun- These results suggest that speech-related rAC/MOF neurons use ing curves, where angles are naturally ordered, vowels can have sparse coding for vowels, in analogy to the sparse bursts in song- different orderings. In the tuning curves of Fig. 2, we ordered birds’ area HVc12 and to the sparse, highly selective responses the vowels according to their place and manner of articulation as observed in the human medial–temporal lobe13. In contradistinc- expressed by their location in the IPA chart1, but is this ordering tion, the gradually modulated speech encoding in STG implies pre- natural to the neural representation? Instead of assuming a cer- viously unrecognized correlates with a hallmark of motor cortical tain ordering, we could try and deduce the natural organization control—broad, sinusoidal tuning, implying a role in motor con- of speech features represented in the population-level neural code. trol of speech production9. Interestingly, speech encoding in these That is, we can try to infer a neighbourhood structure (or order) of anatomical areas is opposite in nature to that of other modalities: the vowels where similar (neighbouring) neuronal representations broad tuning for motor control is common in the frontal lobe10 are used for neighbouring vowels. We reasoned that this neigh- (versus STG in the temporal lobe) whereas, sparse tuning to high- bourhood structure could be extracted from the error structure level concepts is common in the temporal lobe13 (versus rAC/MOF of neuronal classifiers: when a decoder, such as the population in the frontal lobe). Analogous to the recently found subpathway vector11 errs, it is more likely to prefer a value that is a neighbour of between the visual dorsal and ventral streams14, our findings may the correct value than a more distant one. Thus, classification error lend support to a speech-related ‘dorsal stream’ where sensorimotor rates are expected to be higher between neighbours than between prediction supports speech production by a state feedback control3. distant features when feature ordering accurately reflects the neural The sparse rAC/MOF representation may serve as predictor state, representation neighbourhood structure. In this case, classification in line with anterior cingulate15 and orbitofrontal16 roles in reward error rates expressed by the classifier’s confusion matrix will have a prediction. The broad STG tuning may support evidence that the band-diagonal structure. motor system is capable of modulating the perception system to To apply this strategy, we decoded the population firing pat- some degree3,17,18. terns using multivariate linear classifiers with a sparsity constraint Our finding of sharply tuned neurons in rAC/MOF agrees with to infer the uttered vowel (Methods). The five vowels were decoded the DIVA model of the human speech system19, which suggests with a high average (cross-validated) accuracy of 93% (significantly that higher-level prefrontal cortex regions involved in phonologi- above the 20% chance level, P < 10 − 5, one-sided one-sample t-test, cal encoding of an intended utterance sequentially activate speech nature communications | 3:1015 | DOI: 10.1038/ncomms1995 | www.nature.com/naturecommunications © 2012 Macmillan Publishers Limited. All rights reserved. ARTICLE nature communications | DOI: 10.1038/ncomms1995

a c

] 100 80 60 40 20

Decoding accuracy [% 0 i u 2 3 4 5 No. of vowels o

e b Decoded a i e a o u i e a Uttered 100% o u 0 All units

Figure 3 | Inferring the organization of vowel representation by decoding. (a) Average decoding accuracy ( ± s.e.) versus the number of decoded vowel classes. Red dashed lines represent the chance level. (b) Confusion matrix for decoding population activity of all analysed units to infer the uttered vowels. Band-diagonal structure indicates adjacency of vowels in the order a, e, i, u and o in the neural representation. High confusion in the corner (between u and i) implies cyclicity. (c) The vowels IPA chart, representing the highest point of the tongue during articulation, on top of a vocal tract diagram. The inferred connections (blue lines) demonstrate neuronal representation of articulatory physiology. sound map neurons that correspond to the syllables to be produced. (Neuroport, Blackrock). Sorted units (WaveClus30 and SUMU31) recorded in dif- Activation of these neurons leads to the readout of feedforward ferent sessions are treated as different in this study. All studies conformed with the guidelines of the Medical Institutional Review Board at the University of California motor commands to the primary motor cortex. Owing to orbito Los Angeles. frontal connections to both STG20 and ventral pre-motor cortex21, rAC/MOF neurons may participate also in the feedback control Experimental paradigms. Patients first listened to isolated auditory cues (beeps) map, where sharply tuned neurons may provide a high-level ‘dis- and to another individual uttering the vowel sounds and three syllables (me/lu/ha) crete’ representation of the sound to utter, based on STG input from following beeps (auditory controls). Then, following an oral instruction, patients uttered the instructed syllable multiple times, each following a randomly spaced the auditory error map, before low-level commands are sent to the (2–3 s) beep. Syllables consisted of either monophthongal vowels (a/a/, e/ε/, i/i/, articulator velocity and position maps via ventral pre-motor cor- o/o/ and u/u/) or a consonant (of: d/d/, g/g/, h/h/, j/dε/, l/l/, m/m/, n/n/, r/ /, tex. Our broadly tuned STG units may also be part of the transition s/s/ and v/v/), and one of the aforementioned vowels (for example, da/da/, de/dε/, 1 from the auditory error map to the feedback control map, providing di/di/, do/do/ and du/du/) . For simplicity, this paper employs the English rather than IPA transcription as described above. All sessions were conducted at the a lower-level ‘continuous’ population representation of the sound patient’s quiet bed-side. to utter. Our results further demonstrate that the neuronal population Data analysis. Of the 716 recorded units, we analysed 606 that were not respon- encoding of vowel generation appears to be organized according sive during any auditory control (rAC and adjacent MOF cortex (rAC/MOF): 123 to a functional representation of a spatial-anatomical axis: tongue of 156; dorsal and subcollosal ACC: 68/72; entorhinal cortex: 124/138; hippo height. This axis was shown to have a significant main effect on campus: 103/114; amygdala: 92/106; parahippocampal gyrus: 64/66; and STG: 22 32/64). The anatomical subdivisions of the ACC are according to McCormick decoding from speech motor cortex units . Whether these struc- et al.32 Owing to clinical considerations29, no electrode was placed in the primary tured multi-level encoding schemes also exist in other speech areas or pre-motor cortex in this patient population. Each brain region was recorded like Broca’s2 and speech motor cortex, and how they contribute to in at least three subjects. A unit is considered speech-related when the firing rate the coordinated production of speech are important open ques- differs significantly between the baseline ([− 1000, 0]ms relative to the beep) and the response ([0, 200]ms relative to speech onset; paired t-test, P < 0.05, adjusted tions. Notwithstanding, the structured encoding observed naturally for false discovery rate33 control for multiple units and vowels, q < 0.05, n ranges facilitates high-fidelity decoding of volitional speech segments and between 6 and 12 trials depending on the session). For these units, we found the may have implications for restoring speech faculties in individuals maximal response among the four 100-ms bins starting 100 ms before speech who are completely paralysed or ‘locked-in’.23–28 onset, and computed mean firing rates in a 300-ms window around this bin. Tuned units are speech-related units for which the mean firing rate is significantly different between the five vowel groups (analysis of variance;F -test, P < 0.05, false Methods discovery rate33 adjusted, q < 0.05, n between 6 and 12 for each group). Broad, Patients and electophysiology. Eleven patients with pharmacologically resistant sinusoidally tuned units are tuned units whose firing rate correlates with:a + bcos epilepsy undergoing invasive monitoring with intracranial depth electrodes to iden- (c + i2π/5) (where i = 0,…,4 is the index of the vowel in a, e, i, u, o) with coefficient tify the seizure focus for potential surgical treatment29 (9 right handed, 7 females, of determination R2>0.7.10 Sharply tuned units are tuned units for which the mean ages 19–53) participated in a total of 14 recording sessions, each on a different firing rate in the three vowel groups of lowest mean firing rate is the same with day. Based exclusively on clinical criteria, each patient had 8–12 electrodes for 1–2 high probability (analysis of variance; F-test, P > 0.1, n between 6 and 12 for each weeks, each of which terminated with a set of nine 40-µm platinum–iridium group). The vowel decoder is a regularized multivariate linear solver, which mini- mizes ||x|| subject to ||Ax − b|| ≤σ (basis pursuit denoising problem34). It has microwires. Their locations were verified by magnetic resonance imaging or by L1 L2 computer tomography coregistered to preoperative magnetic resonance imaging. superior decoding performance and speed relative to neuron-dropping decoders35. Bandpass-filtered signals (0.3–3 kHz) from these microwires and the sound track A contains the feature inputs to the decoder: spike counts of all units in a baseline were synchronously recorded at 30 kHz using a 128-channel acquisition system bin ([ − 1000,0]ms relative to the beep) and in two 100-ms response bins that fol-

nature communications | 3:1015 | DOI: 10.1038/ncomms1995 | www.nature.com/naturecommunications © 2012 Macmillan Publishers Limited. All rights reserved. nature communications | DOI: 10.1038/ncomms1995 ARTICLE lowed speech onset; b are 5-element binary vectors coding the individual vowels 19. Guenther, F. H. & Vladusich, T. A neural theory of speech acquisition and uniquely. All decoding results were sixfold cross-validated using trials that were production. J. Neurolinguistics 25, 408–422 (2012). not used for decoder training. The decoder was trained on all of the aforemen- 20. Cavada, C., Compañy, T., Tejedor, J., Cruz-Rizzolo, R. J. & Reinoso-Suárez, F. tioned features from the training set only, with no selection of the input neurons The anatomical connections of the macaque monkey orbitofrontal cortex. or their features. Instead, the sparse decoder described above automatically A review. Cereb. Cortex 10, 220–242 (2000). selects task-relevant features by higher weights it allocates to them using 21. Carmichael, S. T. & Price, J. L. Sensory and premotor connections of the orbital the minimal ||x|| constraint; task-unrelated features are thus diminished by and medial prefrontal cortex of macaque monkeys. J. Comp. Neurol. 363, L1 low weights. Owing to the high accuracy in decoding, we randomly dropped 642–664 (1995). 20% of the units (in each cross-validation training) when computing confusion 22. Brumberg, J. S., Wright, E. J., Andreasen, D. S., Guenther, F. H. & Kennedy, P. R. matrices, to increase the amount of confusions and allow the extraction of a Classification of intended phoneme production from chronic intracortical meaningful band-diagonal structure (except for the STG-only training, microelectrode recordings in speech-motor cortex. Front Neurosci. 5, 65 (2011). Supplementary Fig. S4). 23. Guenther, F. H. et al. A wireless brain-machine interface for real-time speech The vowels inFig. 3c were placed on the IPA chart according to the locations synthesis. PLoS ONE 4, e8218 (2009). previously calculated for American speakers (ref. 1, page 42), and the overlaid 24. Brumberg, J. S., Nieto-Castanon, A., Kennedy, P. R. & Guenther, F. H. Brain- connections (blue lines) were inferred by the maximal non-diagonal element computer interfaces for speech communication. Speech Commun. 52, 367–379 for each row and each column of the confusion matrix. (2010). 25. Pasley, B. N. et al. Reconstructing speech from human auditory cortex. PLoS Biol. 10, e1001251 (2012). References 26. Pei, X., Barbour, D. L., Leuthardt, E. C. & Schalk, G. Decoding vowels and 1. International Phonetic Association. Handbook of the International Phonetic consonants in spoken and imagined words using electrocorticographic signals Association (Cambridge University Press, 1999). in humans. J. Neural Eng. 8, 046028 (2011). 2. Sahin, N. T., Pinker, S., Cash, S. S., Schomer, D. & Halgren, E. Sequential 27. Leuthardt, E. C. et al. Using the electrocorticographic speech network to control processing of lexical, grammatical, and phonological information within a brain-computer interface in humans. J. Neural Eng. 8, 036004 (2011). Broca’s area. Science 326, 445–449 (2009). 28. Formisano, E., De Martino, F., Bonte, M. & Goebel, R. ‘Who’ is saying ‘what’? 3. Hickok, G., Houde, J. & Rong, F. Sensorimotor integration in speech Brain-based decoding of human voice and speech. Science 322, 970–973 (2008). processing: computational basis and neural organization. Neuron 69, 29. Fried, I. et al. Cerebral microdialysis combined with single-neuron and 407–422 (2011). electroencephalographic recording in neurosurgical patients. J. Neurosurg. 4. Ghosh, S. S., Tourville, J. A. & Guenther, F. H. A neuroimaging study of 91, 697–705 (1999). premotor lateralization and cerebellar involvement in the production of 30. Quian-Quiroga, R., Nadasdy, Z. & Ben-Shaul, Y. Unsupervised spike detection phonemes and syllables. J. Speech Lang. Hear. Res. 51, 1183–1202 (2008). and sorting with wavelets and superparamagnetic clustering. Neural Comput. 5. Halgren, E. et al. Spatiotemporal stages in face and word-processing. 2. Depth- 16, 1661–1687 (2004). recorded potentials in the human frontal and rolandic cortices. J. Physiol. Paris 31. Tankus, A., Yeshurun, Y. & Fried, I. An automatic measure for classifying 88, 51–80 (1994). clusters of suspected spikes into single cells versus multiunits. J. Neural Eng. 6. Paus, T., Petrides, M., Evans, A. C. & Meyer, E. Role of the human anterior 6, 056001 (2009). cingulate cortex in the control of oculomotor, manual, and speech responses: 32. McCormick, L. M. et al. Anterior cingulate cortex: an MRI-based parcellation a positron emission tomography study. J. Neurophysiol. 70, 453–469 (1993). method. NeuroImage 32, 1167–1175 (2006). 7. Sabri, M. et al. Attentional and linguistic interactions in speech perception. 33. Benjamini, Y. & Hochberg, Y. Controlling the false discovery rate: a practical Neuroimage 39, 1444–1456 (2008). and powerful approach to multiple testing. J. R. Stat. Soc. Ser. B Methodological 8. Buchsbaum, B. R., Hickok, G. & Humphries, C. Role of left posterior superior 57, 289–300 (1995). temporal gyrus in phonological processing for speech perception and 34. van den Berg, E. & Friedlander, M. P. Probing the pareto frontier for basis production. Cognitive Sci. 25, 663–678 (2001). pursuit solutions. SIAM J. Sci. Comput. 31, 890–912 (2008). 9. Ojemann, G. A., Creutzfeldt, O., Lettich, E. & Haglund, M. M. Neuronal 35. Tankus, A., Fried, I. & Shoham, S. Sparse decoding of multiple spike trains for activity in human lateral temporal cortex related to short-term verbal memory, brain–machine interfaces. J. Neural Eng. 9 (in press) (2012). naming and reading. Brain 111 (Pt 6), 1383–1403 (1988). 10. Georgopoulos, A. P., Kalaska, J. F., Caminiti, R. & Massey, J. T. On the relations between the direction of two-dimensional arm movements and cell discharge Acknowledgements in primate motor cortex. J. Neurosci. 2, 1527–1537 (1982). We thank D. Pourshaban, E. Behnke, T. Fields and Prof. P. Keating of UCLA, and Prof. 11. Georgopoulos, A. P., Schwartz, A. B. & Kettner, R. E. Neuronal population A. Cohen and A. Alfassy of the Technion for assistance, and the European Research coding of movement direction. Science 233, 1416–1419 (1986). Council (STG 211055), NINDS, Dana Foundation, Lady Davis and L. and L. Richmond 12. Hahnloser, R. H. R., Kozhevnikov, A. A. & Fee, M. S. An ultra-sparse code research funds for financial support. underlies the generation of neural sequences in a songbird. Nature 419, 65–70 (2002). Author contributions 13. Quian-Quiroga, R., Reddy, L., Kreiman, G., Koch, C. & Fried, I. Invariant visual A.T., I.F. and S.S. designed the study and wrote the manuscript; I.F. performed neuro representation by single neurons in the human brain. Nature 435, 1102–1107 surgeries; A.T. prepared experimental setup and performed the experiments; A.T. and (2005). S.S. analysed the data. 14. Tankus, A. & Fried, I. Visuomotor coordination and motor representation by human temporal lobe neurons. J. Cogn. Neurosci. 24, 600–610 (2011). 15. Hayden, B. Y., Heilbronner, S. R., Pearson, J. M. & Platt, M. L. Surprise Additional information signals in anterior cingulate cortex: neuronal encoding of unsigned reward Supplementary Information accompanies this paper at http://www.nature.com/ prediction errors driving adjustment in behavior. J. Neurosci. 31, 4178–4187 naturecommunications (2011). 16. Wallis, J. D. Cross-species studies of orbitofrontal cortex and value-based Competing financial interests: The authors declare no competing financial interests. decision-making. Nat. Neurosci. 15, 13–19 (2012). Reprints and permission information is available online at http://npg.nature.com/ 17. Hickok, G. & Poeppel, D. The cortical organization of speech processing. reprintsandpermissions/ Nat. Rev. Neurosci. 8, 393–402 (2007). 18. Hickok, G. Computational neuroanatomy of speech production. Nat. Rev. How to cite this article: Tankus, A. et al. Structured neuronal encoding and decoding of Neurosci. 13, 135–145 (2012). human speech features. Nat. Commun. 3:1015 doi: 10.1038/ncomms1995 (2012).