Soma

Soma is my Android companion project. The ambition is to make the phone useful as an interface to speech, models and tools without pretending that connecting those pieces has made them one finished assistant.

A wake word, a transcript and an action are different things. The detector can recognise a phrase while the transcriber gets the next sentence wrong. A transcript can be correct while the response goes to the wrong place. Calling all of that a voice feature makes the feature sound simpler and the failure harder to locate.

Give each problem a boundary

The Android application and the model-training repository are separate. App permissions, service lifecycles and device interactions belong to the mobile side. Training scripts, evaluations and model assets belong to a different workflow, with heavier work handled on appropriate machines.

I want those boundaries to remain visible as the project develops. A model does not gain permission to do something merely because it has produced a plausible instruction. Equally, a neatly drawn architecture does not prove the phone can keep a service alive through the conditions in which it will actually be used.

That is the character of the work at present: connecting capabilities while keeping their limitations inspectable. This page describes an active project, not a catalogue of promises already delivered.

The related speech and wake-word notes are collected below. They deal with the individual pieces, where the useful details and inconvenient qualifications belong.