Skip to content

ADR-v2-008 — True Code-Switch Dictation (make [polyglot] real)

Status: Accepted (2026-07-02) · Wave B Context links: [[adr-v2-000-interaction-layer]], existing [polyglot] stub + polyglot/lid.py

Context

The voice-HCI research (internal) shows monolingual ASR WER spikes 30–50% at code-switch points; end-to-end multilingual models cut boundary WER up to ~55%, and Google Rambler / Mistral Voxtral now ship mid-sentence code-switching — but only in cloud/mobile. There is no offline hold-to-talk equivalent. YazSes already has a [polyglot] config stub and polyglot/lid.py (language-pair parsing + span LID scaffolding), currently dormant.

Decision

Make code-switch dictation real for a user-declared language pair (e.g. fa-en): 1. Segment the burst and run per-span language ID (polyglot/lid.py), constrained to the declared pair to bound LID error. 2. Decode each contiguous same-language run with faster-whisper's language set accordingly (faster-whisper is per-segment), then stitch with existing spacing rules. 3. When the CS-adapted model (adapter_path, trained out-of-band) is present, prefer it; otherwise fall back to per-span monolingual decode.

Config (existing): [polyglot] enabled=false, pair, adapter_path, lid, mer_gate. Wire lid.py span routing into the decode path; keep dormant until pair is set.

Consequences

  • + Natural bilingual dictation, offline — no cloud equivalent exists; reuses the stub.
  • + Bounded LID error via the declared pair.
  • − Mid-utterance LID mistakes cause wrong-language garble → restrict to the declared pair, expose mer_gate; keep off by default.
  • − Per-span re-decode adds latency → only on the polyglot path (monolingual users unaffected).