Deep learning architectures for estimating breathing signal and respiratory parameters from speech recordings

Area of Science:

• Speech Science
• Biomedical Engineering
• Artificial Intelligence

Background:

• Respiration is fundamental to speech production, involving inhalation and exhalation.
• The relationship between speech and respiration is complex, with air outflow modulated by linguistic and prosodic factors.
• Sensing respiratory dynamics directly from speech is plausible but underexplored.

Purpose of the Study:

• To comprehensively investigate techniques for sensing breathing signals and parameters from speech.
• To explore the application of deep learning architectures for respiratory monitoring via speech.
• To address challenges in establishing the practical utility of speech-based respiratory sensing.

Main Methods:

• Utilized deep learning architectures to analyze speech signals.
• Focused on extracting breathing patterns and respiratory parameters from vocalizations.
• Investigated the relationship between speech production and underlying respiratory dynamics.

Main Results:

• Demonstrated the feasibility of sensing breathing signals from speech using deep learning.
• Identified key speech characteristics correlated with respiratory patterns.
• Addressed technical challenges in developing practical speech-based respiratory monitoring.

Conclusions:

• Speech analysis offers a promising non-invasive method for monitoring respiratory dynamics.
• Deep learning models can effectively estimate breathing patterns and parameters from speech.
• This technology has potential applications in understanding and assessing respiratory health.