Why six enrollment phrases? Voiced frames, legacy profile versions and capture mode
Phrase count, voiced estimates and legacy compatibility are separate enrollment conditions
Current enrollment requires 6 phrases and 32000 voiced frames per phrase, unlike the 16000-frame matching floor. voicedRange estimates counts, not total recording seconds or semantic truth. Current v2 profiles require six groups; legacy v1 allows three only with dimension/finite/consistency conditions. This is not arbitrary three-phrase import, mode bypass or identity authentication.
Different from matching, near-field policy and model assets
Matching checks convenient filtering, near-field checks capture policy, assets check fixed files. Here enrollment completeness, profile versions and mode prerequisites matter. No recording or real profile read/replacement ran.
| Check | Verified scope | Manual review | Not established |
|---|---|---|---|
| Current enrollment | enrollmentPhraseCount=6 | Quiet/everyday prompts versus material | Six seconds completes enrollment |
| Per phrase | enrollmentMinimumVoicedSamples=32000 | At 16k about two voiced-estimate seconds, not wall time | Two seconds recorded always enough |
| Matching | minimumVoicedSamples=16000 | Separate stage | Floors interchangeable |
| Incremental consistency | enrollmentIsConsistent can inspect subsets | Partial pass versus completeness | One/two phrases finishes |
| Version | Current v2 six/legacy v1 three | Format/dimensions/finite/consistency | Every old profile compatible |
| File selection | Existing current chosen first | Actual loading outcome | Broken v2 automatically restores v1 |
| Mode | profileCaptureModeMatches prerequisite | Capture toggle versus other routes | Changing mode always retains compatibility |
| Authority | Local enrollment/similarity policy | Actual authority separately | Biometric/spoof/accuracy certification |
Synthetic enrollment note
Three vectors passing a helper do not complete current six-phrase enrollment. Legacy paths still require version/dimension/consistency. Capture-mode changes require original-setting review, not automatic profile replacement or importing another person’s data.
Voiced counts/similarity are algorithmic conditions, not content/identity or real-scene scores. No biometrics/person/device test ran; cross-environment/distance/noise performance is not promised.
FAQ
Are six phrases six total seconds?
No. Phrase count and per-phrase voiced conditions differ.
Do two recorded seconds meet 32000 voiced frames?
Not necessarily; wall duration and voicedRange estimate differ.
Are all old three-phrase profiles accepted?
No. The v1 branch retains other checks, while current v2 needs six groups.
Does partial consistency complete enrollment?
No. Helpers and final completeness differ.
Will broken current profiles automatically restore legacy?
Not guaranteed; file existence selects the path rather than bypassing failed current validation.
Were real profiles read/re-enrolled?
No. Source conditions only, without biometrics or identity claims.
Reading and source
Matching versus authority · Capture policy · Model evidence · Start · Privacy
