Neural Noise

fNIRS and brain responses to AI-synthesized familiar voices

2 min read

  • An fNIRS experimental study examined the impact of AI-synthesized familiar voices on brain neural responses (Nature).1
  • The work is relevant to auditory/social perception, sensory neuroprosthetics, and voice interfaces.1 1

Weekly enrichment (2026-07-20)

  • The study appeared in Scientific Reports (vol. 15, article 16872), accepted 3 March 2025 and published 15 May 2025 (DOI 10.1038/s41598-025-92702-5).23
  • Twenty-five student volunteers aged 20–25 (12 male, 13 female) were recruited, all with normal hearing, no neurological disorders, and a Montreal Cognitive Assessment (MoCA) score ≥ 28; the protocol was approved by the Shaoxing People’s Hospital Ethics Review Committee.4
  • Voices were generated with the open-source GPT-SoVITS model: each participant’s mother read a ~1-minute reference text, and two “stranger” voices (a middle-aged woman and a sweet young-female voice) were added, all in Mandarin, with stimuli text drawn from Long Yingtai’s The Letters of André.4
  • fNIRS was recorded with a NirSmart device (Danyang Huichuang Medical Equipment Co., China) at 760 nm and 850 nm, 11 Hz sampling, and 3 cm probe spacing, deriving oxygenated (HbO) and deoxygenated (HbR) hemoglobin via the modified Beer–Lambert law over prefrontal and temporal cortex.4
  • A block design alternated 25 s voice-listening tasks with 15 s rest across 10 cycles; Experiment 1 (n = 20) compared the maternal voice against a middle-aged-female voice, and Experiment 2 (n = 13) compared it against a sweet-female voice.4
  • In Experiment 1, mean ΔHbO was 0.008703 for the maternal voice versus −0.01329 for the female voice (difference 0.02199), with region-specific differences of 0.01968 (frontal lobe) and 0.02431 (temporal lobe).4
  • In Experiment 2, mean ΔHbO was 0.01641 for the maternal voice versus −0.006584 for the sweet-female voice (difference 0.02300), with differences of 0.01946 (frontal lobe) and 0.02653 (temporal lobe).4
  • A linear mixed-effects model (fixed effects: voice type, brain region, and their interaction; random effect: participant) confirmed the AI-synthesized maternal voice significantly increased prefrontal and temporal activation, pointing to familiarity processing tied to emotion and memory that is relevant to affective voice interfaces and neuroprosthetic feedback.24

Footnotes

  1. https://news.google.com/rss/articles/CBMiX0FVX3lxTFBBeTdIRkJPS3ptWmtrUFFpVmhWMmJlZm1haUZFRzFKN0VLYlR3MEhubFZla05TU3BrMHkyWlZCbTFkSnAtR1RuQTR6TFRhRFZGV1NjODNrSjZHM0xJZjdv?oc=5 ↩ ↩2 ↩3

  2. https://www.nature.com/articles/s41598-025-92702-5 ↩ ↩2

  3. https://ora.ox.ac.uk/objects/uuid:fbffa7e2-ca8c-461b-9866-50706a7ead91 ↩

  4. https://pmc.ncbi.nlm.nih.gov/articles/PMC12081930/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

Graph View