A recorded memory, a therapy session, a parent playing with a child: each carries pauses, prosody and word choices that say something about memory and mood, and no clinician can score them at scale. How much of that signal can be captured from French clinical audio, compared over time, and shared without exposing the speaker? The team will build the pieces: who spoke when, voice and language markers, and anonymisation that protects identity while keeping the signal. Speech foundation models, language models and on-device processing work on open corpora, and the two days end with a plan to bring them to the clinic. Still wide open: markers that survive a phone call.