research/README.md

Research scripts

How the calibrated models were trained. They need a local copy of the MIT Interview Dataset (not included; see its license).

1. Convert the recordings to 16 kHz mono FLAC (ffmpeg -i P1.wav -ac 1 -ar 16000 P1.flac).
2. python diarize.py separates the candidate from the interviewer (writes diar/*.json).
3. python train_calibration.py extracts features with calibration.py, cross-validates per-trait models and writes assets/models/calibrated_scores.joblib.
4. python benchmark.py is the research benchmark using the dataset's own facial, smile and PRAAT features.

Put turker_scores_full_interview.csv and interview_transcripts_by_turkers.csv from the dataset next to the scripts.