Skip to content

Latest commit

 

History

History
131 lines (94 loc) · 4.44 KB

README.rst

File metadata and controls

131 lines (94 loc) · 4.44 KB

mcd

This package computes mel cepstral distortions in python. Mel cepstral distortions are used in assessing the quality of synthesized speech.

Overview

Mel cepstral distortion (MCD) is a measure of how different two sequences of mel cepstra are. It is used in assessing the quality of parametric speech synthesis systems, including statistical parametric speech synthesis systems, the idea being that the smaller the MCD between synthesized and natural mel cepstral sequences, the closer the synthetic speech is to reproducing natural speech. It is by no means a perfect metric for assessing the quality of synthetic speech, but is often a useful indicator in conjunction with other metrics.

The mcd package provides scripts to compute a variety of forms of MCD score:

  • plain MCD, for which it is assumed that the two sequences to be compared are already "aligned" in terms of their timing.
  • plain MCD excluding certain segments, for example silence segments.
  • MCD DTW, which uses dynamic time warping (DTW) to compute the minimum MCD obtainable by "aligning" the two sequences. This metric does not penalize differences in the timing between natural and synthetic speech, which is often desirable.

It also contains general purpose dynamic time warping code.

License

Please see the file License for details of the license and warranty for mcd.

Installation

For most purposes the simplest way to install mcd is to use pip. For example in Debian and Ubuntu:

sudo apt-get install python-numpy
sudo pip install mcd

The first command installs numpy from the system repository, since installing numpy using pip is generally not recommended. The second command installs the latest released version of mcd on PyPI, together with any currently uninstalled python packages required by mcd.

mcd can also be installed in a virtualenv:

sudo apt-get install python-numpy
virtualenv --system-site-packages env
env/bin/pip install mcd

The latest development version of mcd is available from a github repository (see below).

To check that mcd is installed correctly you can run the test suite:

python -m unittest discover mcd

Examples

Examples of example usage (in unix) are given in example_usage.

Development

The source code is hosted in the mcd github repository. To obtain the latest source code using git:

git clone git://github.com/MattShannon/mcd.git

Development is in fact done using darcs, with the darcs repository converted to a git repository using darcs-to-git.

To install any currently uninstalled python packages required by mcd:

sudo apt-get install cython python-numpy
sudo pip install -r requirements.txt

To compile the cython part of mcd in the current directory:

python setup.py build_ext --inplace

This command must be run after every modification to the source .pyx files.

To run the full test suite, including tests of command-line tools, on the working copy:

python -m unittest discover mcd
PYTHONPATH=. python bin/test_cli.py

A note on setup.py

The included setup.py file operates in one of two modes depending on whether or not the file dev is present in the project root directory. In development mode (dev present, as for the github repository), the build_ext command uses cython to compile cython modules from their .pyx source, and the sdist command is modified to first use cython to compile cython modules from their .pyx source to .c files. In distribution mode (dev absent, as for source distributions such as the code on PyPI), the build_ext command uses a C compiler to directly compile cython modules from the corresponding .c files. This approach ensures that source distributions can be installed on systems without cython or with an incompatible version of cython, while ensuring that distributed .c files are always up-to-date and that the source .pyx files are used instead of .c files during development.

The author would welcome any suggestions for more elegant ways to achieve a similar effect to the approach described above!

Bugs

Please use the issue tracker to submit bug reports.

Contact

The author of mcd is Matt Shannon.