deceptive-speech

ML for classifying deceptive speech from audio/text

Data

Original data is labeled transcribed sequences of Deceptive/Not Deceptive speech audio from a (few hundred?) people
Data is broken into 3 types: Acoustic, Lexical, and Personal
Acoustic and Lexical are at two granularities: IPU level, and Turn level
- Acoustic: OpenSMILE (IS13) acoustic/prosodic feature extractor
- Lexical: DAL lexical sentiment feature extractor
- Lexical: LIWC lexical sentiment feature extractor
Personal is collected per person.
- These are NEO-FFI personality indicators: Neurotiscism, Extraversion, Openness, Agreeableness, and Conscientiousness
- Also the persons gender
- And whether their native language is Mandarin Chinese or American English (1/0?)

Given train / dev/ test IPU splits:

Split	Num IPUs	%	Num People	%
Train	19390	35	60	37
Dev	12386	22	51	31
Test	23766	43	51	31

** Will combine training and dev and do cross validation to have roughly 60/40 split **

Naive baseline accuracy: 62.12% train, 59.04% test

Name		Name	Last commit message	Last commit date
Latest commit History 17 Commits
Paper		Paper
data		data
.gitignore		.gitignore
Baselines.ipynb		Baselines.ipynb
Big IPU-Turns Data Cleaning.ipynb		Big IPU-Turns Data Cleaning.ipynb
Data Cleaning.ipynb		Data Cleaning.ipynb
IPUs.ipynb		IPUs.ipynb
Initial Exploration.ipynb		Initial Exploration.ipynb
README.md		README.md
python_requirements.txt		python_requirements.txt
turns-big-train-small-test.ipynb		turns-big-train-small-test.ipynb
turns-no-is13.ipynb		turns-no-is13.ipynb
turns-small-train-big-test.ipynb		turns-small-train-big-test.ipynb
turns.ipynb		turns.ipynb
turns_small_all.ipynb		turns_small_all.ipynb
turns_small_no_praat.ipynb		turns_small_no_praat.ipynb