Lars Hausfeld

Task-Dependent Decoding of Speaker and Vowel Identity from Auditory Cortical Response Patterns

Selective attention to relevant sound properties is essential for everyday listening situations. It enables the formation of different perceptual representations of the same acoustic input and is at the basis of flexible and goal-dependent behavior. Here, we investigated the role of the human auditory cortex in forming behavior-dependent representations of sounds. We used single-trial fMRI and analyzed cortical responses collected while subjects listened to the same speech sounds (vowels /a/, /i/, and /u/) spoken by different speakers (boy, girl, male) and performed a delayed-match-to-sample task on either speech sound or speaker identity. Univariate analyses showed a task-specific activation increase in the right superior temporal gyrus/sulcus (STG/STS) during speaker categorization and in the right posterior temporal cortex during vowel categorization. Beyond regional differences in activation levels, multivariate classification of single trial responses demonstrated that the success with which single speakers and vowels can be decoded from auditory cortical activation patterns depends on task demands and subjects behavioral performance. Speaker/vowel classification relied on distinct but overlapping regions across the (right) mid-anterior STG/STS (speakers) and bilateral mid-posterior STG/STS (vowels), as well as the superior temporal plane including Heschls gyrus/sulcus. The task dependency of speaker/vowel classification demonstrates that the informative fMRI response patterns reflect the top-down enhancement of behaviorally relevant sound representations. Furthermore, our findings suggest that successful selection, processing, and retention of task-relevant sound properties relies on the joint encoding of information across early and higher-order regions of the auditory cortex.

Brain-Based Translation: fMRI Decoding of Spoken Words in Bilinguals Reveals Language-Independent Semantic Representations in Anterior Temporal Lobe

Bilinguals derive the same semantic concepts from equivalent, but acoustically different, words in their first and second languages. The neural mechanisms underlying the representation of language-independent concepts in the brain remain unclear. Here, we measured fMRI in human bilingual listeners and reveal that response patterns to individual spoken nouns in one language (e.g., “horse” in English) accurately predict the response patterns to equivalent nouns in the other language (e.g., “paard” in Dutch). Stimuli were four monosyllabic words in both languages, all from the category of “animal” nouns. For each word, pronunciations from three different speakers were included, allowing the investigation of speaker-independent representations of individual words. We used multivariate classifiers and a searchlight method to map the informative fMRI response patterns that enable decoding spoken words within languages (within-language discrimination) and across languages (across-language generalization). Response patterns discriminative of spoken words within language were distributed in multiple cortical regions, reflecting the complexity of the neural networks recruited during speech and language processing. Response patterns discriminative of spoken words across language were limited to localized clusters in the left anterior temporal lobe, the left angular gyrus and the posterior bank of the left postcentral gyrus, the right posterior superior temporal sulcus/superior temporal gyrus, the right medial anterior temporal lobe, the right anterior insula, and bilateral occipital cortex. These results corroborate the existence of “hub” regions organizing semantic-conceptual knowledge in abstract form at the fine-grained level of within semantic category discriminations.

EEG decoding of spoken words in bilingual listeners: from words to language invariant semantic-conceptual representations.

Spoken word recognition and production require fast transformations between acoustic, phonological, and conceptual neural representations. Bilinguals perform these transformations in native and non-native languages, deriving unified semantic concepts from equivalent, but acoustically different words. Here we exploit this capacity of bilinguals to investigate input invariant semantic representations in the brain. We acquired EEG data while Dutch subjects, highly proficient in English listened to four monosyllabic and acoustically distinct animal words in both languages (e.g., “paard”–“horse”). Multivariate pattern analysis (MVPA) was applied to identify EEG response patterns that discriminate between individual words within one language (within-language discrimination) and generalize meaning across two languages (across-language generalization). Furthermore, employing two EEG feature selection approaches, we assessed the contribution of temporal and oscillatory EEG features to our classification results. MVPA revealed that within-language discrimination was possible in a broad time-window (~50–620 ms) after word onset probably reflecting acoustic-phonetic and semantic-conceptual differences between the words. Most interestingly, significant across-language generalization was possible around 550–600 ms, suggesting the activation of common semantic-conceptual representations from the Dutch and English nouns. Both types of classification, showed a strong contribution of oscillations below 12 Hz, indicating the importance of low frequency oscillations in the neural representation of individual words and concepts. This study demonstrates the feasibility of MVPA to decode individual spoken words from EEG responses and to assess the spectro-temporal dynamics of their language invariant semantic-conceptual representations. We discuss how this method and results could be relevant to track the neural mechanisms underlying conceptual encoding in comprehension and production.

Pattern analysis of EEG responses to speech and voice: Influence of feature grouping

Pattern recognition algorithms are becoming increasingly used in functional neuroimaging. These algorithms exploit information contained in temporal, spatial, or spatio-temporal patterns of independent variables (features) to detect subtle but reliable differences between brain responses to external stimuli or internal brain states. When applied to the analysis of electroencephalography (EEG) or magnetoencephalography (MEG) data, a choice needs to be made on how the input features to the algorithm are obtained from the signal amplitudes measured at the various channels. In this article, we consider six types of pattern analyses deriving from the combination of three types of feature selection in the temporal domain (predefined windows, shifting window, whole trial) with two approaches to handle the channel dimension (channel wise, multi-channel). We combined these different types of analyses with a Gaussian Naïve Bayes classifier and analyzed a multi-subject EEG data set from a study aimed at understanding the task dependence of the cortical mechanisms for encoding speakers identity and speech content (vowels) from short speech utterances (Bonte, Valente, & Formisano, 2009). Outcomes of the analyses showed that different grouping of available features helps highlighting complementary (i.e. temporal, topographic) aspects of information content in the data. A shifting window/multi-channel approach proved especially valuable in tracing both the early build up of neural information reflecting speaker or vowel identity and the late and task-dependent maintenance of relevant information reflecting the performance of a working memory task. Because it exploits the high temporal resolution of EEG (and MEG), such a shifting window approach with sequential multi-channel classifications seems the most appropriate choice for tracing the temporal profile of neural information processing.

The effect of spatial resolution on decoding accuracy in fMRI multivariate pattern analysis.

Multivariate pattern analysis (MVPA) in fMRI has been used to extract information from distributed cortical activation patterns, which may go undetected in conventional univariate analysis. However, little is known about the physical and physiological underpinnings of MVPA in fMRI as well as about the effect of spatial smoothing on its performance. Several studies have addressed these issues, but their investigation was limited to the visual cortex at 3T with conflicting results. Here, we used ultra-high field (7T) fMRI to investigate the effect of spatial resolution and smoothing on decoding of speech content (vowels) and speaker identity from auditory cortical responses. To that end, we acquired high-resolution (1.1mm isotropic) fMRI data and additionally reconstructed them at 2.2 and 3.3mm in-plane spatial resolutions from the original k-space data. Furthermore, the data at each resolution were spatially smoothed with different 3D Gaussian kernel sizes (i.e. no smoothing or 1.1, 2.2, 3.3, 4.4, or 8.8mm kernels). For all spatial resolutions and smoothing kernels, we demonstrate the feasibility of decoding speech content (vowel) and speaker identity at 7T using support vector machine (SVM) MVPA. In addition, we found that high spatial frequencies are informative for vowel decoding and that the relative contribution of high and low spatial frequencies is different across the two decoding tasks. Moderate smoothing (up to 2.2mm) improved the accuracies for both decoding of vowels and speakers, possibly due to reduction of noise (e.g. residual motion artifacts or instrument noise) while still preserving information at high spatial frequency. In summary, our results show that - even with the same stimuli and within the same brain areas - the optimal spatial resolution for MVPA in fMRI depends on the specific decoding task of interest.

Hearing an Illusory Vowel in Noise: Suppression of Auditory Cortical Activity

Human hearing is constructive. For example, when a voice is partially replaced by an extraneous sound (e.g., on the telephone due to a transmission problem), the auditory system may restore the missing portion so that the voice can be perceived as continuous (Miller and Licklider, 1950; for review, see Bregman, 1990; Warren, 1999). The neural mechanisms underlying this continuity illusion have been studied mostly with schematic stimuli (e.g., simple tones) and are still a matter of debate (for review, see Petkov and Sutter, 2011). The goal of the present study was to elucidate how these mechanisms operate under more natural conditions. Using psychophysics and electroencephalography (EEG), we assessed simultaneously the perceived continuity of a human vowel sound through interrupting noise and the concurrent neural activity. We found that vowel continuity illusions were accompanied by a suppression of the 4 Hz EEG power in auditory cortex (AC) that was evoked by the vowel interruption. This suppression was stronger than the suppression accompanying continuity illusions of a simple tone. Finally, continuity perception and 4 Hz power depended on the intactness of the sound that preceded the vowel (i.e., the auditory context). These findings show that a natural sound may be restored during noise due to the suppression of 4 Hz AC activity evoked early during the noise. This mechanism may attenuate sudden pitch changes, adapt the resistance of the auditory system to extraneous sounds across auditory scenes, and provide a useful model for assisted hearing devices.

Learning of new sound categories shapes neural response patterns in human auditory cortex

The formation of new sound categories is fundamental to everyday goal-directed behavior. Categorization requires the abstraction of discrete classes from continuous physical features as required by context and task. Electrophysiology in animals has shown that learning to categorize novel sounds alters their spatiotemporal neural representation at the level of early auditory cortex. However, functional magnetic resonance imaging (fMRI) studies so far did not yield insight into the effects of category learning on sound representations in human auditory cortex. This may be due to the use of overlearned speech-like categories and fMRI subtraction paradigms, leading to insufficient sensitivity to distinguish the responses to learning-induced, novel sound categories. Here, we used fMRI pattern analysis to investigate changes in human auditory cortical response patterns induced by category learning. We created complex novel sound categories and analyzed distributed activation patterns during passive listening to a sound continuum before and after category learning. We show that only after training, sound categories could be successfully decoded from early auditory areas and that learning-induced pattern changes were specific to the category-distinctive sound feature (i.e., pitch). Notably, the similarity between fMRI response patterns for the sound continuum mirrored the sigmoid shape of the behavioral category identification function. Our results indicate that perceptual representations of novel sound categories emerge from neural changes at early levels of the human auditory processing hierarchy.

Horizontal tuning for faces originates in high-level Fusiform Face Area

Recent work indicates that the specialization of face visual perception relies on the privileged processing of horizontal angles of facial information. This suggests that stimulus properties assumed to be fully resolved in primary visual cortex (V1; e.g., orientation) in fact determine human vision until high-level stages of processing. To address this hypothesis, the present fMRI study explored the orientation sensitivity of V1 and high-level face-specialized ventral regions such as the Occipital Face Area (OFA) and Fusiform Face Area (FFA) to different angles of face information. Participants viewed face images filtered to retain information at horizontal, vertical or oblique angles. Filtered images were viewed upright, inverted and (phase-)scrambled. FFA responded most strongly to the horizontal range of upright face information; its activation pattern reliably separated horizontal from oblique ranges, but only when faces were upright. Moreover, activation patterns induced in the right FFA and the OFA by upright and inverted faces could only be separated based on horizontal information. This indicates that the specialized processing of upright face information in the OFA and FFA essentially relies on the encoding of horizontal facial cues. This pattern was not passively inherited from V1, which was found to respond less strongly to horizontal than other orientations likely due to adaptive whitening. Moreover, we found that orientation decoding accuracy in V1 was impaired for stimuli containing no meaningful shape. By showing that primary coding in V1 is influenced by high-order stimulus structure and that high-level processing is tuned to selective ranges of primary information, the present work suggests that primary and high-level levels of the visual system interact in order to modulate the processing of certain ranges of primary information depending on their relevance with respect to the stimulus and task at hand.

The Effect of Task Instruction on Haptic Texture Processing: The Neural Underpinning of Roughness and Spatial Density Perception

Perceived roughness is associated with a variety of physical factors and multiple peripheral afferent types. The current study investigated whether this complexity of the mapping between physical and perceptual space is reflected at the cortical level. In an integrative psychophysical and imaging approach, we used dot pattern stimuli for which previous studies reported a simple linear relationship of interdot spacing and perceived spatial density and a more complex function of perceived roughness. Thus, by using both a roughness and a spatial estimation task, the physical and perceived stimulus characteristics could be dissociated, with the spatial density task controlling for the processing of low-level sensory aspects. Multivoxel pattern analysis was used to investigate which brain regions hold information indicative of the level of the perceived texture characteristics. While information about differences in perceived roughness was primarily available in higher-order cortices, that is, the operculo-insular cortex and a ventral visual cortex region, information about perceived spatial density could already be derived from early somatosensory and visual regions. This result indicates that cortical processing reflects the different complexities of the evaluated haptic texture dimensions. Furthermore, this study is to our knowledge the first to show a contribution of the visual cortex to tactile roughness perception.

Multiclass fMRI data decoding and visualization using supervised self-organizing maps

When multivariate pattern decoding is applied to fMRI studies entailing more than two experimental conditions, a most common approach is to transform the multiclass classification problem into a series of binary problems. Furthermore, for decoding analyses, classification accuracy is often the only outcome reported although the topology of activation patterns in the high-dimensional features space may provide additional insights into underlying brain representations. Here we propose to decode and visualize voxel patterns of fMRI datasets consisting of multiple conditions with a supervised variant of self-organizing maps (SSOMs). Using simulations and real fMRI data, we evaluated the performance of our SSOM-based approach. Specifically, the analysis of simulated fMRI data with varying signal-to-noise and contrast-to-noise ratio suggested that SSOMs perform better than a k-nearest-neighbor classifier for medium and large numbers of features (i.e. 250 to 1000 or more voxels) and similar to support vector machines (SVMs) for small and medium numbers of features (i.e. 100 to 600voxels). However, for a larger number of features (>800voxels), SSOMs performed worse than SVMs. When applied to a challenging 3-class fMRI classification problem with datasets collected to examine the neural representation of three human voices at individual speaker level, the SSOM-based algorithm was able to decode speaker identity from auditory cortical activation patterns. Classification performances were similar between SSOMs and other decoding algorithms; however, the ability to visualize decoding models and underlying data topology of SSOMs promotes a more comprehensive understanding of classification outcomes. We further illustrated this visualization ability of SSOMs with a re-analysis of a dataset examining the representation of visual categories in the ventral visual cortex (Haxby et al., 2001). This analysis showed that SSOMs could retrieve and visualize topography and neighborhood relations of the brain representation of eight visual categories. We conclude that SSOMs are particularly suited for decoding datasets consisting of more than two classes and are optimally combined with approaches that reduce the number of voxels used for classification (e.g. region-of-interest or searchlight approaches).