The lung speaks out-and machines are listening

Listening to audible sound for medical diagnosis and monitoring spans health care professionals, physiologists, psychologists, and computer engineers. Artificial intelligence has advanced acoustic analysis in both voice/speech science and respiratory medicine, but asymmetrically. This asymmetry has deep roots: McKusick's adoption of sound spectrography and Forgacs' oscilloscope-based waveform inspection applied methods established in voice/speech acoustics. The same one-way pattern persists: sound-classification methods migrate from voice/speech research into lung sound analysis with a multi-year lag.

I argue this transfer carried speech science's engineering strengths while omitting an achievement of classical respiratory acoustics: analysis referenced to airflow. Flow-gated phonopneumography established that lung sounds must be interpreted relative to the mechanical act producing them, a validated principle that is absent from contemporary AI-supported auscultation. I trace this omission from Laennec through 19th-century mechanical/acoustic conflations to self-supervised learning, showing how label ontology and missing physiological referencing constrain what these systems can learn, explain, or teach.

Restoring flow-referenced analysis is not a return to outdated methodology but a correction to a field that adopted a neighboring discipline's instrumentation without its foundational insight. Parallels-annotator-quality gaps and physiology-referenced efforts, where voice/speech science is neither ahead nor behind-suggest that respiratory and voice/speech acoustics could collaborate to mutual benefit. Lung sound AI's next generation should be built on physiological grounding rather than acoustic pattern-matching alone, a principle also for all professionals who use their auditory perception in assisting health care and for the computer engineers building the tools that listen alongside them.