working on the problem that a transcript doesn't know how you said something. models turn audio into plain text and everything carried by volume, speed, and emphasis dies in the conversion. the plan is a layer between my voice and the text that stamps that meaning back in.
the prototype is live. loudness and speed already show up as prosody annotations in square brackets, and now the emotion side is calibrating with a small model that classifies tone. there's a calibration page where i record takes for laugh, sigh, sarcasm, question, and excited, and the pipeline scores each one.
it fought back in dumb ways. two frontends serve the calibration page and one had a hardcoded list of takes that never learned about the new recordings, plus an old ssl change had silently broken its private proxy to the backend. fixed both, verified end to end, forty five takes logged through the app.
the studio page grew dry run and apply buttons, so calibration is now self serve: record, analyze offline, apply the thresholds when they look right. the event detector still has zero positive training examples, so laugh and sigh recordings are next.