Speech Recognition via fNIRS Based Brain Signals.

Area of Science:

• Neuroimaging
• Brain-Computer Interfaces
• Speech Perception

Background:

• Non-invasive neuroimaging techniques are crucial for understanding brain activity in real-world settings.
• Functional-near infrared spectroscopy (fNIRS) offers portability and wearability for ecological studies.
• Decoding perceived speech from brain signals is a key challenge in neuroscience.

Purpose of the Study:

• To demonstrate the feasibility of identifying perceived speech using fNIRS brain signals.
• To develop and evaluate machine learning models for speech decoding based on hemodynamic responses.
• To explore the potential of fNIRS in creating speech-decoding brain-computer interfaces.

Main Methods:

• Participants listened to audio stories while brain activity was monitored using fNIRS.
• Machine learning models were trained on fNIRS data to predict listened story segments.
• Generalized Canonical Correlation Analysis (GCCA) was used to address inter-subject variability in fNIRS signals.

Main Results:

• Accurate differentiation of story segments was achieved using fNIRS data.
• An average accuracy of 74.7% was obtained for distinguishing between two 50-second story segments.
• An average accuracy of 43.6% was achieved for differentiating four 25-second story segments.

Conclusions:

• fNIRS-based brain signal analysis can identify perceived speech.
• The study highlights the potential of fNIRS for developing novel speech decoding brain-computer interfaces.
• This approach could lead to new types of neural prosthetic systems.