Implemented a simple RNN lie detector on the CSC Deceptive Speech Dataset using TensorFlow 1.7 Estimator API.
I tried modeling the data in two different ways; first (per frame method), pass in a MFCC vector with each row as a time step to the RNN, for instance, if the MFCC vector has dimension (28000,13), then the input for the RNN model is (28000, batch_size, 13); second , pass in the MFCC vector such that the whole vector is treated as the same time step, i.e. (1, 1, 28000 * 13). The two method can be selected using flag --per_frame.
The the MFCC vectors have different length, in order to pass in the vectors into TensorFlow's Estimator API, I had to add paddings to the vectors that are shorter than the longest vector, and use the sequence_length parameter in dynamic_rnn to ignore the padding. The model support the basic RNN, LSTM and GRU which can be specified with flag --model. When using the per frame method, the last output state from dynamic_rnn is then taken to apply dropout with flag --dropout, and pass to the output layer.
python3 lie-detector.py --lr=0.001 ---epochs=5 --model=LSTM --data_dir=data --per_frame=True