Clone this repo, then run:

# Set up virtualenv
virtualenv . && source bin/activate

# Install package requirements
pip install -r requirements.txt

To build the training data (with prepare.py), you'll need the following packages:

• SPHERE, for decoding from the OGI sound formats
• SoX, for sound preprocessing
• openSMILE, for audio feature extraction

The binaries from these pacakages need to be on your PATH for the corpus handling scripts to work properly.

We use the OGI Multilanguage Corpus for training and evaluation. You can acquire a copy of this corpus through the LDC.

With a fresh download of the OGI corpus, run prepare.py to get the corpus into a format usable with this software:

python prepare.py <OGI_PATH> ge,ma

In this command the argument ge,ma specifies that we'd like to perform language detection with German and Mandarin recordings.

The prepare.py script generates a directory with devtest, evaltest, and train Pickle files for each provided language. To train on this data, run the command

python nodules.py -d <data_dir> train

The nodules.py script will train SVM and logistic regression models with default settings and output them to a local directory. Training can take anywhere from 5 minutes (baseline) to 24 hours (SVM, complex feature set, lots of data).

You can use the same nodules.py script to evaluate trained models. The train phase should save model files with .jbl extensions. Provide such a file to the nodules.py script like so:

python nodules.py -d <data_dir> --model-in-file <model_file> -v test