Location Spotting Using Prosody
Installing and Running
Training and evaluation is done through the bidirectional_keras.py file Python 3 and GPU are required to run code as is.
The code requires that you use a GPU, for the LSTM.
To make CPU compatible: the Tensorflow.keras.layers call of CuDNNLSTM should be changed to Tensorflow.keras.layers.LSTM, This will make the code run much slower on GPU
Required Python packages
- tensorflow-gpu (1.13 tested)
Training and Evaluating
Depending on what the code is being used for, there are 4 parameters that should be adjusted
train mode if is True, then the model will be trained.
single_track if is True, then only the prosodic features of the user speaking will be used.
balanced_eval if is True, then the model will evaluated on a balanced dataset where half of the frames are location frames, and the other half are frames with speech that don't have a location mention.
eval_other_lang if is True, then the model will be evaluated over data from another language, as specified in the mat_path variable
Authors for the research paper are: Gerardo Cervantes Nigel G. Ward