NAOMS · building block

Speech

Speech recognition and text-to-speech for talking to your app.

What a person notices

Talking to the app works offline, with audio processed on the device and deleted right after.
Nothing is sent anywhere else for processing unless the person has explicitly agreed to it.
A voice conversation is searchable afterward, the same way a typed note would be.

Speech turns voice into text and text into voice, with the local path as the default and the remote path opt-in only.

Two paths exist side by side: a fully local path with on-device speech-to-text and text-to-speech and no network calls at all, and a remote path for heavier processing that requires an explicit, checked consent grant before any audio is touched.
On the local path, audio files are deleted immediately after transcription. Nothing lingers waiting to be cleaned up later.
Transcriptions are stored automatically as memories in the personal knowledge graph, so a conversation held by voice becomes searchable the same way a typed note would.
The remote path checks that an audio file actually lives in an allowed location before touching it, and filters out low-confidence transcription segments rather than passing along a guess as if it were certain.