Swiss Dialect Identification

In this project I compare different machine learning setups for the identification of 4 Swiss German dialects.

Setup

For the setup of this repository simply type:

make

This will

• set up a virtual environment for this repository,
• install all necessary project dependencies.

If anything does not work, try installing with Python 3.8.

Clean and Re-install

To reset the repository to its inital state, type:

make dist-clean

This will remove the virtual environment and all dependencies.
With the make command you can re-install them.

To remove temporary files like .pyc or .pyo files, type:

make clean

Data

The data for this project comes from 3rd VarDial Evaluation Campaign from 2019 (for more information see: https://sites.google.com/view/vardial2019/campaign).

It consists of a training and a validation set.
The data comprises utterances from 4 Swiss German dialects (Basel, Bern, Lucerne, Zurich). The dataset comes with three different representations of the utterances.

• The files train.txt and dev.txt contain an utterance per line combined with its dialect label. The label is separated from the utterance by a tab.
• The files train.vec and dev.vec contain a so-called iVector per line. The iVector is a vector of floating-point numbers, which represents the spoken form of the utterances. The iVector has been extracted from audio files.
• (The original dataset contained another file for training and test set that kept the normalized forms of the utterances. This data is not needed for this project and cannot be found in the repository.)

The training set contains 14279 utterances and the development set contains 4530 utterances.

Models and Training

Models

• Multinomial Logistic Regression (Baseline Model)
• SVM

Training

It is possible to train a single model (-s), perform cross validation (-v) to select a best model or train a final model (-f) on the whole data set.
Besides, there is the chance to use the original train-test split (-o).
Now it is also possible to include GAN-generated data into the training process (-g). A file name must be given as an argument.
Eventually, you can choose to train with print output or not (-v).

Examples:

Train a single SVM model with original split:

python3 scripts/train.py -s SVM -o

Train a single logistic regression model with randomized split (implementation still missing):

python3 scripts/train.py -s LogisticRegression

Choose a best model with cross validation:

python3 scripts/train.py -c

Train the best model (here assumed it is an SVM) on the whole dataset (train and dev):

python3 scripts/train.py -f SVM

NOTE: Only a single model (not final) can be trained on the original split.

Storing a Model

Models can be stored in the repository.
At the end of each training run you are asked whether you want to store the output model or not.

Models are stored in the directory stored_models.

Data Manipulation

The experiments with the data in this project have shown a strong influence of the underlying speaker representation on dialect identification.
Different approaches have been taken to manipulate the data in such a way that a stronger emphasis is on the dialect differences.

CGAN

A Conditional Generative Adversarial has been and can be trained to create new data, which creates ivectors (representing pseudo-utterances) from assumed pseudo-speakers that do not belong to the original dataset.

There are already CGAN models, which can be used for ivector generation:

python3 scripts/cgan.py -g file_name.pt

A new model can also be trained:

python3 scripts/cgan.py -t

It is also possible to train a new model and generate from it directly:

python3 scripts/cgan.py -t -g file_name.pt

After training a prompt will ask whether the model should be stored.
Models are stored in the directory stored_models.

Once configuration files are supported, the settings for the GAN can be changed through the configuration file.
At the moment values will have to be changed directly in the script.

Configuration Files

The functionality is not implemented yet but the aim is to store training parameters in configuration files. There will also be a script to generate configuration files.

Logging

Train losses and metrics are stored in the directory train_logs.
All information about each training run can be found here.