Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Contributed Talk Session: Thursday, August 14, 11:00 am – 12:00 pm, Room C1.03
Poster Session C: Friday, August 15, 2:00 – 5:00 pm, de Brug & E‑Hall

Disentangling of Spoken Words and Talker Identity in Human Auditory Cortex

Akhil Bandreddi¹, Dana Boebinger¹, David Skrill¹, Kirill V Nourski², Matthew Howard, Christopher Garcia, Thomas Wychowski, Webster Pilcher, Samuel Victor Norman-Haignere³; ¹University of Rochester, ²The University of Iowa, ³University of Rochester Medical Center

Presenter: Akhil Bandreddi

Complex natural sounds such as speech contain many different types of information, but recognizing these distinct information sources is computationally challenging because sounds with shared information differ widely in their acoustics. For example, variation across talkers makes it challenging to recognize the identity of a word, while variation in the acoustics of different words makes it challenging to recognize talker identity. How does the human auditory cortex disentangle word identity from talker identity such that each type of information can be coded invariant to acoustic variation all other information sources? To address this question, we measured neural responses to a diverse set of 338 words spoken by 32 different talkers using spatiotemporally precise intracranial recordings from the human auditory cortex. We developed a simple set of model-free experimental metrics for quantifying representational disentangling of word and talker identity, both within individual electrodes as well as across different dimensions of the neural population response. We observed individual electrodes that show a representation of words that is partially robust to acoustic variation in talker identity, but no electrodes or brain regions showed a robust representation of talker identity. However, at the population level, we observed distinct dimensions of the neural response that nearly exclusively reflected either words or talker identity, and were completely invariant to acoustic variation in the non-target dimension. These results suggest that while there is partial specialization for talker-robust word identity in localized brain regions, robust disentangling is accomplished at the population level with distinct representations of words and talker identity mapped to distinct dimensions of the neural code for speech.

Topic Area: Language & Communication

Extended Abstract: Full Text PDF