feat: offline pronunciation pipeline with NeMo CTC + GOP scoring

[fasuizu-br]

Summary

• Adds fully offline, privacy-first pronunciation evaluation using NeMo Conformer CTC Small (INT4, 17.1MB) via ONNX Runtime Web
• Per-phoneme GOP scoring with Viterbi forced alignment — identifies exactly which sounds the student mispronounces
• L1-aware Brazilian Portuguese accent adaptation (21 phonemes across 3 tiers) with decoded transcript verification
• Replaces the Web Speech API approach from PR feat: add speaking practice with Web Speech API #19 which auto-corrected student speech (useless for pronunciation evaluation)

Technical details

Metric	Value
Model size	17.1 MB (INT4, 61% smaller than FP32)
WER (native)	2.9%
WER (BR accent)	17.8% (does NOT auto-correct)
Unit tests	155 passing
Cross-browser	Chrome, Firefox, Safari (incl. iOS)
Privacy	100% offline — no audio leaves device

Pipeline

AudioWorklet → WebWorker → ONNX inference → Viterbi alignment → GOP scoring → L1 adaptation

L1 Scoring Tiers (Brazilian Portuguese)

• Tier 1 (50% boost): TH, DH, R, NG, ZH — absent in Portuguese
• Tier 2 (40% boost): AE, IH, AH, UH, EY, OW, ER, AY, AW, OY — vowel confusion
• Tier 3 (25% boost): L, T, D, S, Z — context-dependent differences

Real Audio Validation (Speech Accent Archive, George Mason University)

• 12 Brazilian speakers + 6 native Americans
• 50% BR-to-native gap recovered by L1 scoring
• Statistical significance: p=0.0017, Cohen's d=1.28 (large effect)

Files changed

• src/lib/speechUtils.ts — CTC processing, Viterbi alignment, GOP scoring, L1 adaptation
• src/workers/stt-worker.ts — WebWorker for ONNX inference
• src/hooks/useSpeechRecognition.ts — React hook (audio capture + worker communication)
• src/components/Study/SpeakingPractice.tsx — UI with per-phoneme colored feedback
• src/lib/types.ts — TypeScript types for pronunciation results
• public/models/ — INT4 ONNX model (17.1MB) + token vocabulary

Test plan

• npm run build passes cleanly
• npx vitest run — 155 tests passing
• Manual test: record pronunciation, verify per-phoneme scores appear
• Verify L1 feedback tooltips show BR-specific phoneme adjustments
• Test on mobile (Android Chrome, iOS Safari)
• Verify RAM usage < 80MB in Chrome DevTools