Abstract 501, Date 8:00 am - 12:00 pm, Monday, February 8, 2010|
Session Session M:
|Spectro-Temporal Encoding of Speech by ECoG Signals in Human Auditory and Motor Cortices|
|*Stephen David, Brian Pasley, Nima Mesgarani, Adeen Flinker, Edward Chang, Nathan Crone, Robert Knight, Shihab A. Shamma|
Research in non-human mammals has described the spectro-temporal representation of speech and other natural sounds in cortex, but little is known about how these findings generalize to humans. To study auditory representation in humans, we recorded electrocorticographic signals (ECoG) from epileptic patients during a 24 hour period. Subdural recordings were made using a grid of electrodes (spacing 10 mm) placed over temporal and frontal lobes. Patients were presented with a sequence of speech sounds (isolated words and sentences) during passive listening.
To characterize auditory tuning, spectro-temporal receptive fields (STRFs) were estimated using the time-varying high gamma power (100-300 Hz, 10 ms time bins) in the ECoG signal at each recording site. STRFs were estimated by normalized reverse correlation, a procedure that compensates for correlations present in speech and other natural sounds that can bias STRFs estimated using other methods.
Sites along the lateral surface of the superior temporal gyrus showed clear tuning to sound features, and the corresponding STRFs were able to predict high gamma activity in a validation data set with correlations up to r=0.4. STRFs measured from power at lower ECoG frequencies and the raw signal showed weaker tuning; high-gamma STRFs performed consistently better. Preliminary topographic data revealed systematic variability in bandwidth across sites. For a small number of sites in primary motor cortex, STRFs also had significant predictive power, suggesting that these areas participate in processing the basic features of speech, even during passive listening. A large number of sites in temporal cortex that could not be characterized with STRFs did show phase-locked responses to auditory stimuli. Characterization of tuning at these sites may be possible with nonlinear spectro-temporal models or models that incorporate high-level abstract sound features.