Raw Speech Classification

Trains CNN (or any neural network based) classifiers from raw speech using Keras and tests them. The inputs are lists of wav files, where each file is labelled. It then creates fixed length signals and processes them. During testing, it computes scores at the utterance or speaker levels by averaging the corresponding frame-level scores from the fixed length signals.

Installation

From source, In a Conda environment

To install Keras 3 with PyTorch backend, run:

conda env create -f conda/rsclf-pytorch.yaml

To install Keras 3 with TensorFlow backend, run:

conda env create -f conda/rsclf-tensorflow.yaml

Then install the package in that environment (the default name is rsclf) with:

conda run -n rsclf pip install .

Installing from PyPI

If you want to install the last release of this package in your current environment, you can run either of the following commands depending on your desired framework:

pip install raw-speech-classification[torch]

or

pip install raw-speech-classification[tensorflow]

You'll also need to set the KERAS_BACKEND environment variable to the correct backend before running rsclf-train or rsclf-test (see below), or globally for the current bash session with:

export KERAS_BACKEND=torch

Replace torch by tensorflow accordingly.

Using the code

1. Create lists for training, cross-validation and testing. Each line in a list must contain the path to a wav file (relative to the -R or --root option), followed by its integer label indexed from 0, separated by a space. E.g. if your data files are in /home/bob/data/my_dataset/part*/file*.wav, the root option could be /home/bob/data/my_dataset and the content of the files would then be like:

   part1/file1.wav 1
   part1/file2.wav 0
   

   Full list files for IEMOCAP are available in the repository as example in datasets/IEMOCAP/F1_lists.

2. If you installed from source with Conda: A run script is available that concatenates all the steps. Run run.sh. Provide the model architecture as an argument. See model_architecture.py for valid options. Optionally, provide an integer as a count of the number of times the experiment is repeated. This is useful when the same experiment needs to be repeated multiple times with different initialization. The argument defaults to 1.

   If you installed with pip: You can run the following commands (give the --help option to each command for more details):

   rsclf-wav2feat --wav-list-file list_files/cv.list --feature-dir output/cv_feat --mode train --root path/to/dataset/basedir
   rsclf-wav2feat --wav-list-file list_files/train.list --feature-dir output/train_feat --mode train --root path/to/dataset/basedir
   rsclf-wav2feat --wav-list-file list_files/test.list --feature-dir output/test_feat --mode test --root path/to/dataset/basedir
   KERAS_BACKEND=torch rsclf-train --train-feature-dir output/train_feat --validation-feature-dir output/cv_feat --output-dir output/cnn_subseg --arch subseg --splice-size 25 --verbose 2
   KERAS_BACKEND=torch rsclf-test --feature-dir output/test_feat --model-filename output/cnn_subseg/cnn.keras --output-dir output/cnn_subseg --splice-size 25 --verbose 0
   rsclf-plot --output-dir output/ output/cnn_subseg
   

This is an example of how to run on the IEMOCAP dataset using conda, assuming conda is installed in ~/miniconda3 and your environment is rsclf:

bash run.sh -C ~/miniconda3 -n rsclf -D ./datasets/IEMOCAP/F1_lists -a seg -o results/seg-f1 -R <IEMOCAP_ROOT>

For instance, <IEMOCAP_ROOT> can be /ssd/data/IEMOCAP which should contain IEMOCAP_full_release/Session*.

This is an example of the log printed to the terminal, and you should obtain the following curve in results/seg-f1/plot.png:

Code components

1. wav2feat.py creates directories where the wav files are stored as fixed length frames for faster access during training and testing.

2. train.py is the Keras training script.

3. Model architecture can be configured in model_architecture.py.

4. rawdataset.py provides an object that reads the saved directories in batches and retrieves mini-batches for training.

5. test.py performs the testing and generates scores as posterior probabilities. If you need the results per speaker, configure it accordingly (see the script for details). The default output format is:

    <speakerID> <label> [<posterior_probability_vector>]
   

6. plot.py generates and saves the learning curves.

Training schedule

The script uses stochastic gradient descent with 0.5 momentum. It starts with a learning rate of 0.1 for a minimum of 5 epochs. Whenever the validation loss reduces by less than 0.002 between successive epochs, the learning rate is halved. Halving is performed until the learning rate reaches 1e-7.

Contributors

Idiap Research Institute

Authors: S. Pavankumar Dubagunta and Dr. Mathew Magimai-Doss

License

GNU GPL v3