The code in this repository can be used to identify phonetic and phonological patterns in German word-final neutralization using Kaldi ASR. The included experiments test the distribution of incomplete neutralization, complete neutralization, and no neutralization (that is, voiceless, partially voiced, and voiced consonants) in word-final position by using ASR to identify phones in word-final position. The model is trained to assume that underlyingly voiced consonants are always realized as semi-voiced consonants in word-final position. However, the model is able to learn the features of fully voiced and fully voiceless consonants from other contexts. It is then asked to determine the distribution of these phones in the given environments.
Because this system uses bash scripting, it is generally restricted to Unix systems (including Mac OS X), though it may also be possible to get it working on Windows under Cygwin.
The incomplete neutralization experiments use Kaldi ASR for training the model and MATLAB for analyzing it.
- To install Kaldi ASR, follow the instructions at the Kaldi ASR documentation.
- To purchase and install MATLAB, visit the MATLAB page at Mathworks.
It is also necessary to download the VM1 corpus and to unzip the documentation contained within.
Once these have been installed, clone this repository.
$ git clone https://github.com/thorsonlinguistics/german-neutralization.git
This can be put anywhere, but it is easiest to place it within Kaldi's egs
directory.
This will create a directory german-neutralization which contains the
following sub-directories. You do not need to take any action with these.
german-neutralization/conf/- configuration files for Kaldi-ASRgerman-neutralization/steps/- useful scripts from Kaldigerman-neutralization/utils/- useful scripts from Kaldigerman-neutralization/local/- useful local scripts
It will also create the following files:
german-neutralization/config.sh- local configurationgerman-neutralization/path.sh- path variablesgerman-neutralization/run.sh- a script to run the model
It is necessary to modify config.sh and path.sh with paths pointing to your
installation of Kaldi, the VM1 corpus, and the VM1 documentation.