SAFE consists of two modules: training models and F0
estimation. During training, SAFE loads wave files and their F0
ground truth files to statistically estimate models for
noise robust F0 estimation. During F0 estimation, SAFE evaluates
the likelihoods of F0 candidates on each frame given the
trained models, then generates an F0 contour that maximizes
the overall likelihoods.
How SAFE is built
SAFE is a toolkit written in C and built on the
General Scientific Library (GSL),
FFTW library, and another Speech Processing
Assistant (SPA) library, which is developed by Wei Chu and
Prof. Abeer Alwan. SAFE is released under the Less
General Public License (LGPL). Everyone is
encouraged to contribute to SAFE. The current
version is 1.0.
Documentations and FAQs
The package for
downloading has a brief
tutorial explaining
the softwares and packages that are required,
the formats of input and output, and how to run the toolkit
with example scripts. If you still encounter problems, feel
free to let Wei know.
Requirements
So far, we have only compiled and tested the Linux version of SAFE.
We will probably developed a Windows version in the near future.
To run SAFE, you need to have:
• a 32-bits Fedora (≥12) or Ubuntu (≥8) system;
• Gcc (≥4.4);
• Automake (≥ 1.0);
• Autoconf (≥ 2.0);
• Libtool (≥ 2.2);
• General
scientific library (GSL) and its development package (≥
1.14);
• FFTW and
its development package (≥ 3.2)
• Sound eXchange
(SoX) (≥ 14.0)
SAFE is written in
C, you do not need Matlab.
We would like to thank Georg Meyer for providing the KEELE
database; CSTR lab at Edinburg University for providing
another FDA database. The databases were used to evaluation
the algorithm.
Contacts
If you have any questions and suggestion, feel free to let
us know.
Wei Chu (weichu@ucla.edu) is maintaining the toolkit and website.
