Assist naming across transcripts with voice

After a seed cast is named and linked, enrol trusted voice and pre-load suggestions so Speaker Identification can propose the same people on later managed transcripts. You still confirm each suggestion.

Outcome

You will have enabled local voice matching, enrolled reference samples from confirmed profile links, pre-loaded suggestion caches, and seen a suggested profile on a later transcript that you can confirm.

Starting point

  • The speaker_match extra is installed (Models). Default installs do not include it.

  • Managed library transcripts with linked audio (enrol and analyse need speech; short turns under about eight seconds may not match).

  • At least one transcript where you named speakers and created or attached longitudinal profiles (not name-only).

  • Local AI / Ollama is not required.

What you’ll do

  1. Enable voice matching and read the privacy notice.

  2. Enrol trusted voice for all profiles (reference corpus).

  3. Pre-load voice suggestions (sample all managed speakers).

  4. Open Speaker Identification on another transcript and confirm the suggested person.

Walkthrough

  1. Open Settings → Speakers. Read the voice privacy notice and enable voice matching. Consent is stored in privacy.voice_settings.json. Opt-in does not enrol any samples.

  2. Confirm you already have active profiles with confirmed links (from Speaker Identification → Create new profile, or Attach to an existing person). Profiles without links cannot enrol.

  3. Under Library voice batch, choose Refresh enrol inventory, then Enrol trusted voice for all profiles. This walks confirmed links up to Max confirmed links per voice enrol (default 40). Archived and merged profiles are skipped.

  4. When enrol finishes, choose Refresh suggestion inventory, then Pre-load voice suggestions. This analyses every non-ignored, non-collision speaker on managed transcripts into .cache/voice. That is the library-wide “sample all” step. An empty reference corpus still analyses but usually returns No reliable match — enrol first.

  5. Open Speaker Identification on a later managed transcript. For each diarized speaker, the Link to speaker profile control should list a voice or name match when one is available. Review the chip, then save to confirm. Scores never write a name by themselves.

  6. Optionally use Analyse all speakers on one transcript instead of library Pre-load when you only need that file, or after caches go stale.

What to notice

  • Enrol all builds references; Pre-load / Analyse all builds queries. Order matters.

  • Ignored speakers and collision keys are skipped by Pre-load.

  • Suggestion-assisted samples are not trusted references until they are enrolled after you confirm more links. Re-run Enrol all if matches stay weak.

  • Duplicate profiles (two “Maya”s) make matching worse — attach to the existing person instead of creating another.

  • Revoke consent or Delete voice evidence wipes samples; the enrol-cap setting survives revoke.

You should now have…

A reference corpus for linked people, warmed suggestion caches, and a later transcript where Speaker Identification proposes a person you can confirm.

Next