Speaker diarization answers “who spoke when?”
Diarization analyzes a single recording and groups regions of speech into temporary labels such as Speaker 1 and Speaker 2. It does not need to know anyone’s real name. Its job is to distinguish voices within the conversation and associate the right transcript segments with each one.
The useful output is a speaker-aware transcript even when every label is temporary.
Voiceprint matching answers “might this be a known person?”
A voiceprint is an embedding: a compact numerical representation drawn from speech. A local library can compare a newly separated speaker’s aggregated audio with previously confirmed samples. This can surface a candidate such as “Possible: Maya.” Similarity scores help rank candidates; they are not a calibrated probability of identity.
It should not turn into an unreviewed identity claim. Audio quality, microphones, languages, overlap, and similar-sounding speakers all affect confidence.
A safer workflow: candidates before entities
ZoneVoice uses a staged model for this reason. After diarization, the app can collect enough clear audio from an unnamed temporary speaker and form a local candidate group. Across later recordings, similar voices can strengthen that group. The user decides when it becomes a named person.
- Separate: create temporary speakers for one recording.
- Aggregate: select clear segments and make a normalized local voiceprint vector.
- Compare: rank new speakers against the local library using similarity.
- Suggest: show possible matches only where confidence and the lead over the next match are strong.
- Confirm: save a user-approved name and retain the match status.
Why the second-best match matters
A single high similarity score can look persuasive on its own. It is more useful when considered alongside the next-best candidate. If two names score almost the same, the system has weak evidence to choose one. In that case the right product behavior is to leave the voice unconfirmed or ask the user to choose.
Why might no candidate appear?
Importing a file alone does not create a voiceprint. Speaker identification must finish, and a separated voice needs enough usable audio. Short, noisy, or overlapping segments can leave too little clear speech to form a candidate.
Check that processing included speaker identification, review whether the transcript has separated speakers, and try a longer recording with clear speech. Leave uncertain identities unnamed rather than forcing a match.
What stays local in ZoneVoice
ZoneVoice is designed to keep voiceprint vectors in local SQLite storage encrypted with AES-256-GCM. Speaker labels, aliases, confirmation state, model version, and match score belong to the local workspace. The desktop app does not require a cloud speaker database to make these suggestions.