Shape the voice / smooth, movable anatomy

Organ outlines are closed subpixel contours extracted from the seven supplied binary masks using a 0.65-pixel Gaussian field and a small contour regularizer. Cubic curves provide a smooth vector display. Tiny components below one pixel of contour area are omitted from that display; original mask edges remain available. Smoothing does not add MRI resolution or establish anatomical accuracy.

Three tongue handles bend the tip, dorsum, and root independently using a smooth interpolating deformation field. The other two handles and three inferior anchors stay in place when a handle moves. The selected handle’s dashed outline shows its observed motion envelope. The jaw rotates around an approximate posterior pivot. Lip motion is local to the anterior tissue; velum and larynx have local vertical controls. The airway walls move with their adjacent organs and are remeasured after editing. These attachment fields are geometric approximations, not a muscle model or a full airspace resegmentation. Control limits are recomputed for each MRI frame from this clip: tongue landmarks use trajectory convex hulls; the other controls use observed lip, jaw, velum, and larynx measurements. Deformed tissue contours are constrained to their mask union over all 210 frames with a 0.85-pixel smoothing allowance (or their original contour distance plus 0.02 pixels, if larger). Tongue landmarks at contour indices 25, 55, and 94 estimate tip, dorsum, and root; the supplied contour has no anatomical landmark labels. These clip-specific limits are not a measured physiological range or proof that every mixed pose is attainable. A local folding guard and sampled airway-contact guard further limit edits; arbitrary tissue collisions elsewhere are not solved.

Editing holds the current MRI frame and switches to synthesis. Recording mode plays the attached speech and uses its playback clock to select MRI frames. Both files default to the same start time; their duration difference is about 5.25 ms, which does not independently prove acquisition synchronization. The alignment offset permits adjustment: positive values show a later MRI frame for the same audio time. Controls never alter the reference recording.

The physical sound model retains a persistent 44-section waveguide and a smooth LF-shaped glottal derivative source with aspiration, gradual coefficient changes, and output filtering. During synthesis playback, the optional source follower uses estimated pitch, periodicity, and amplitude from the attached WAV. Held-frame editing sustains the manual pitch. These estimates do not reproduce the speaker’s voice, consonants, or prosody exactly. Shape changes are sent continuously to the audio thread. There is no nasal side branch, so moving the velum changes the modeled oral gap but does not reproduce nasalization. No speaker reconstruction model has been trained on the recording. A valid candidate source-to-lips path exists in 113/210 original frames; disconnected frames remain visible and are muted in synthesis. The recorded speech is unaffected.