MMSE Regression Pipeline (Spec-Compliant)

Implements the user-provided formulation:

• Task: supervised regression y_hat_MMSE = f(x_acoustic, x_text)
• Acoustic features: MFCC (mean/std/delta/delta2), F0 stats, speech rate, pause ratio, mean pause duration, energy (librosa).
• Linguistic features: TTR, MLU, POS noun ratio, disfluency rate, bigram log-prob, BERT [CLS] embedding (768d).
• Models: Ridge (grid alpha), SVR (RBF), GradientBoostingRegressor. StandardScaler inside pipeline. CV: KFold=5, seed=42.
• Metric: RMSE (mean/std).

Structure

• feature_extraction/acoustic.py — librosa features per spec.
• feature_extraction/linguistic.py — spaCy + transformers features per spec.
• train.py — build features, train Ridge/SVR/GBR, select best by RMSE, save bundle to models/.
• predict.py — load bundle, extract features, emit JSON with predicted_mmse, model_used, rmse, feature_importance.
• config.json — paths and model names.

Usage

1. CSV đầu vào:
   • Ưu tiên: có audio_path, transcript, mmse.
   • Nếu không có audio/transcript, script sẽ fallback dùng toàn bộ cột số (trừ id/target) với target là một trong mmse/mms/mmse2.
2. Cài deps: pip install -r requirements.txt (cần librosa, spacy, transformers, torch; tải spaCy model: python -m spacy download en_core_web_sm).
3. Train: python train.py --config config.json
4. Predict: python predict.py --config config.json --audio /path/audio.wav --transcript "..."
   Nếu model được train trên feature sẵn có (không có audio_path), bạn cần cung cấp vec đặc trưng tương ứng thay vì audio+transcript (hiện script predict ưu tiên audio+transcript).
   Script in ra JSON kết quả.

Notes

• Features are numeric only; NaN/Inf are replaced with 0.0.
• Predictions are clipped to [0, 30].
• Reproducibility: random seed = 42. CV is subject-independent if the CSV is prepared that way.