From Typing and Speech to Training Data

There is a considerable amount of work between an interaction and a useful training example. The interaction is messy; the training script would prefer a much tidier account of what happened.

The keyboard work separates collection from learning. The recorded design keeps model training out of the live typing path. Structured records capture inputs and outcomes, while offline processing prepares material for later work. A model package being used on the phone is not the same thing as a model updating itself on every keystroke.

Make the stages inspectable

The harvest workflow documents extraction, conversion and archive steps for preparing data, including processing associated with Titan. Each step needs a defined input and output. A directory full of observations is not automatically a dataset, and a dataset is not evidence that a model has been trained or deployed.

Speech work adds its own preparation and evaluation questions. Audio format, segmentation and recording conditions affect what an example can tell me. Synthetic material can be useful without being an adequate substitute for recordings made through the intended microphone.

The privacy requirement belongs at the collection boundary, not in a reassuring paragraph added after the pipeline exists. What is collected, retained and permitted to enter a training set needs to be explicit. This page states that requirement; it does not claim an audit has proved every historical record anonymous or every storage location free of sensitive material.

The pipeline remains development work. Specific completed runs and measured outcomes belong in dated notes with their evidence attached.