Use stock English Whisper for the English→Chin direction

[trinitron88]

Fixes the "street → strih" garbling in English→Chin mode.

Cause

The cnh-fine-tuned model specialized on Hakha Chin and lost English ASR fidelity (catastrophic-forgetting-lite). In en→cnh mode it was used to transcribe English, so it re-spelled English words into Chin phonetics — "street" → "strih" — and the bad token then passed through translation into both transcript lines.

Fix

• english_asr_model() lazily loads a stock English Whisper (EN_ASR_MODEL, default small.en) on first en→cnh use — zero cost if that direction is never used. Falls back to the fine-tuned model if it can't load.
• on_utterance routes en→cnh through the English model; cnh→en keeps the fine-tuned model (its strength, unaffected).
• en→cnh translation already worked (DIRECTIONS["en2cnh"] = sl=en, tl=cnh, line ~341) — no change needed there, confirmed.
• APP_VERSION → v5.2.0-enasr; README documents EN_ASR_MODEL.

Notes

• Safe: lazy load + fallback, so it can't break boot or the cnh→en path.
• Memory: adds a small English model (default small.en, ~244M) on the GPU, only when en→cnh is first used.
• Built on top of the now-merged caching (Implement translation + TTS caching (real code) #15) and source-defined version.

Deploy test: CT_BRANCH=feature/english-asr-en2cnh python hf_space/deploy_colab.py, switch to English → Hakha Chin, say "live on a street" — the source line should read real English now, and the Chin translation should follow.

🤖 Generated with Claude Code