The train/test splits for all 3 datasets used in our evaluation (e.g., Speech Commands v2, VoxForge and Ambient Context) is provided in the corresponding directories. The splits are provided in the form of .txt files, providing the filepath to the corresponding .wav file.
In the config.yml file, the path to the desired dataset is expected to be provided. A typical top-level directory layout of DATA_DIR is:
. # Root directory of dataset.
├── Data # Directory contraining all dataset samples.
├────Train # Directory contraining all train files.
├────Test # Directory contraining all test files.
├── train_split.txt # File contraining the train .wav samples filepaths with their corresponding labels.
├── test_split.txt # File contraining the test .wav samples filepaths with their corresponding labels.
To prepare the Speech Commnads dataset directory, run the following commands:
foo@bar:~$ mkdir speech_commands && cd speech_commands
foo@bar:~$ wget http://download.tensorflow.org/data/speech_commands_v0.02.tar.gz -O train.tar.gz && mkdir -p ./Data/Train && tar -zxvf ./train.tar.gz -C ./Data/Train
foo@bar:~$ wget http://download.tensorflow.org/data/speech_commands_test_set_v0.02.tar.gz -O test.tar.gz && mkdir -p ./Data/Test && tar -zxvf ./test.tar.gz -C ./Data/Test
The above commands will create the Data directory. In addition to this, you need to copy the corresponding train-test .txt files from data_splits to the root speech_commands directory.