Multilingual Code-Switching Aligner

This offline prototype counts a small Spanish vocabulary in supplied text and suggests an English/Spanish mixing style. It does not identify arbitrary matrix languages, support Hinglish, or establish improvements in inclusivity, trust, or cognitive load.

The helper returns a style label only. A separate host could use it as a suggestion, subject to the caller's stated language preference; rapport or cognitive-load benefits are not established.

Scientific Foundation

Paper / Framework	Relevance
Output Language Alignment for CSW	Demonstrates that AI mirroring code-switching frequencies increases user trust and reduces linguistic anxiety.
Matrix Language Frame (MLF) model	Myers-Scotton framework for distinguishing embedded words vs the matrix language.
FCA Consumer Duty (Vulnerability)	Reduces cognitive load for non-native speakers by allowing them to use their natural hybrid dialects.

How it works

1. Supply a transcript string. An optional host ASR component is outside this contribution.
2. It strips punctuation and detects embedded vocabularies to calculate a Code-Mixing Index (CMI).
3. The system returns a structured CodeSwitchingReport containing the CMI and the suggested PromptStyle.
4. A host may review the suggested style before updating a prompt; the helper performs no prompt or speech mutation.

Decision Matrix

Code-Mixing Index (CMI)	Derived Prompt Style	Expected LLM Output Behavior
CMI == 0.0	monolingual_english	Strict monolingual English (standard).
0.0 < CMI < 0.30	low_code_switching	Occasional embedded loan words (e.g., "gracias", "pero").
CMI >= 0.30	high_code_switching	Fluid Spanglish; alternating sentence clauses.

Expected Outcomes & Metrics

Latency is a design target, not a measured integration benchmark.

Metric	Target	Notes
CMI Computation Latency	< 5ms	Runs via ultra-fast lexicon matching.
Rapport / Trust Score	Not measured	No A/B evaluation of this helper is supplied.

Limitations & Known Constraints

• Lexicon Coverage: The prototype uses a hardcoded vocabulary set. Production systems should use dynamic NLP models to classify embedded words for any language pair.
• ASR Dependency: Requires an ASR model capable of transcribing code-switched audio without forcing translations (e.g. Whisper large-v3).

Integration

This is an offline text helper. Real-time middleware, ASR, language preferences, and LLM prompt updates are proposed host integration work.