Remember named speakers across recordings (auto-identify via speaker embeddings)

Right now every new meeting or transcription starts with generic labels (Speaker 1, Speaker 2…), and I have to rename everyone by hand each time, even people I’ve already named in earlier recordings. The autocomplete of past names helps, but I still have to work out who is who in every recording.

Proposal: when I assign a real name to a speaker, store a voice profile for that person (e.g. the averaged speaker embedding from the diarization pipeline’s VAD/segmentation output). In future recordings, compare each new speaker cluster’s embedding against the saved profiles (e.g. cosine similarity above a threshold) and pre-fill the name automatically, with no input needed for speakers Vowen has seen before.

Nice to have:

• A confidence threshold, with unmatched speakers left as “Speaker N”

• Easy correction: if I rename an auto-labelled speaker, update or split the profile

• A “People / Voice profiles” page to view, merge or delete profiles

• Profiles kept local, consistent with Vowen’s on-device approach

Related: “voice recognition” (auto-label my own voice). This would generalise that to every known speaker.