This is the Kyoto Fst Decoder (Kyfd), a general purpose beam-search decoder for weighted finite state transducers (WFSTs).
It has been used for a number of applications including statistical machine translation, speaking style transformation, OCR correction, grapheme-to-phoneme conversion, and speech segmentation.
Kyfd v. 0.0.5
The source repository can be found at github.
Before installing Kyfd, OpenFst and Xerces-C++ must be installed.
The main executable in the Kyfd package is the Kyfd decoder.
It can be run as follows.
kyfd [options...] config.xml
An example configuration file can be found here, with descriptions of the various options.
Options specified at the command line will override options specified in the configuration file.
These are a few short tutorials on using the Kyfd decoder.
- Simple Modeling/Decoding: Shows how to create a simple FST model to represent a word dictionary, and uses OpenFst to re-discover word segmentations from a corpus with spaces deleted.
- Combining Multiple Models (Under Construction)
- Training Log-Linear Weights (Under Construction)
- Using Hierarchichal Failure Transitions (Under Construction)
Kyfd is currently developed solely by Graham Neubig, but any additional developers are welcome. If you are interested, please send an email to kyfd@.
Kyfd is released under the Apache License, Version 2.0
Kyfd - Release 0.0.5 (2/1/2014)
- Added an option "-sample" that allows you to derive N samples from the output distribution instead of getting the N best.
Kyfd - Release 0.0.4 (6/2/2012)
- Upgraded Kyfd so it works with OpenFst 1.3.2.
Kyfd - Release 0.0.3 (1/19/2012)
- Fixed a few minor bugs.
- Added the ability to print out all symbols with their corresponding alignments using the "-printall" command.
Kyfd - Release 0.0.2 (4/12/2011)
- Updated Kyfd to fix a compile error that occured with newer versions of OpenFST (thanks to Shinichiro Hamada for pointing this out).
Kyfd - Release 0.0.1 (9/2/2009)
- Support for Beam-search or best-first decoding of WFST models
- Storing and output of the contribution of multiple models for later log-linear weight training
- Output of n-best lists
- Support for hierarchical failure transitions