Polyglot Switch — origin note¶

Written: 2026-06-14 · Owner: Mohsen Seyedkazemi Ardebili Shipped as: design/v2-cognitive-layer/04-polyglot-switch.md · ADR-v2-008 Tier: design/ — public. Historical origin note; the implementation docs above are the current source of truth.

The idea¶

The pitch came from lived bilingual experience: real speech doesn't respect language boundaries mid-sentence, but every dictation tool forces a pick-one-language-per-utterance mode. The honest boundary was stated up front — this was never a pitch for universal any-language code-switching (the field can't do that offline yet), but for a single user-configured pair, detecting the seam mid-utterance and transcribing each span in its own orthography.

What shipped¶

Code-switch routing shipped as src/yazses/polyglot/ — a pure language-identification and routing layer that parses a configured pair (e.g. "fa-en"), picks a span's dominant language, and detects within-pair switching. It is off by default. The survey's own boundary held: the routing layer is real and shipped, but the code-switch-adapted transcription model it routes to is trained out-of-band and stays dormant until an adapter path is configured — the feature is a real, working seam waiting on a model, not a finished end-to-end capability for an arbitrary pair.