Objective:
- Using the CMU Deceptive Speech Corpus (CDC), develop a neural network that correctly distinguishes lies from truths in interview recordings.
- Cleaning the data was challenging, as it appeared that there were some areas when the audio transcriptions did not exactly align with the recorded audio.
- A greedy algorithm was used to align these fragments and k-means was used determine a loudness theshold to help strip leading and lagging silence from the audio clips.
- Models mainly consisted of a series of LSTMs; the output of which was combined in different ways.
- We also prototyped a model that used a set of stacked, dilated 1D convolutions over the encoded input, roughly inspired by WaveNet.
- Simpler models performed just as well as more complex models.
- The most important factors for increasing performance was the addition of transcript data encoded in GloVe vectors.
- Previous work on this subject could benefit from (1) better feature selection and (2) more rigorous cross validation techniques for establishing accuracy estimates.
- Previous work only used SVM classifiers on aggregate acoustic measurement features. Essentially, previous work recorded an accuracy that was roughly consistent with the majority class distribution.
- Our work yielded evidence to support that (1) LSTMs can be used effectively on this task, (2) lexical information appears to be more predictive than acoustic features and (3) using a more rigourous rotating, single speaker test set, our test set accuracy was closer to 78.5% instead of the roughly ~63% accuracy that was previously observed.
- To strengthen these conclusions, it would be necessary to investigate the per-speaker distributions of lies vs. truths.
