Voiceprint Mind — origin note¶

Written: 2026-06-14 · Owner: Mohsen Seyedkazemi Ardebili Shipped as: design/v2-cognitive-layer/01-voiceprint-mind.md · ADR-v2-009 Tier: design/ — public. Historical origin note; the implementation docs above are the current source of truth.

The idea¶

The pitch: every dictation tool meets a new user cold, mishearing names, jargon, and accents it has never encountered. The dream was a recogniser that quietly gets more fluent in you the more you talk — with no training UI, no labelling chore, and the resulting personalization living only on the user's own disk, never a vendor's profile.

What shipped¶

The first, cheap layer of this — biasing Whisper's decoding with the user's own frequent vocabulary and personal terms mined from their own dictation history, no model training involved — shipped as src/yazses/personalize/prompt_builder.py, wired to the personalize feature toggle (off by default). The heavier half of the original idea — an opt-in nightly LoRA fine-tune of the acoustic model itself, gated on a held-out WER improvement before ever taking effect — is designed and specified but not yet built; it remains a documented next step rather than a shipped capability.