Reconstructing voice identity from noninvasive auditory cortex recordings

Area of Science:

• Neuroscience
• Auditory Perception
• Machine Learning in Neuroscience

Background:

• Temporal voice areas (TVAs) process conspecific vocalizations in primates.
• Neural representation of speaker identity within TVAs is not well understood.

Purpose of the Study:

• To investigate how speaker identity is represented in the brain.
• To explore the utility of a deep neural network-derived voice latent space (VLS) for understanding neural representations of voice identity.

Main Methods:

• Utilized a deep neural network (DNN) to create a voice latent space (VLS).
• Employed encoding, representational similarity, and decoding analyses to relate VLS to fMRI data.
• Analyzed responses to thousands of voice stimuli from hundreds of speakers.

Main Results:

• The VLS linearly mapped onto fMRI activity in temporal voice areas (TVAs).
• VLS better explained representational geometry for speaker identity in TVAs compared to primary auditory cortex (A1).
• TVA-based reconstructions using VLS preserved speaker identity.

Conclusions:

• The DNN-derived VLS provides high-level representations of voice identity in TVAs.
• This study advances understanding of neural coding for speaker recognition.