Husky2

Husky2 is the first fully pipeleline based speech decoder. The pipeleline based design enables highly abstracted implementations of speech recognition systems while keeping the access to the details of the decoding process, which makes Huksy2 suitable for prototyping and education. In Husky2, everything is viewd as data flow just like the traditional signal processing. It is a WFST based decoder written by the Haskell pure functional programming language using the lazy (or delayed) evaluation.

husky software package download (sourceforge)
A sample speech model for husky
(Please download this file to run test.sh in the husky software package. The file size is 42Mbyte)
husky2.1 reference manual
husky development note How to install haskell compiler, introductions to haskell etc. (in Japanese)
sample input files with various length

*Related paper:
Takahiro Shinozaki, Sadaoki Furui, Yasuo Horiuchi and Shingo Kuroiwa "Pipeline Decomposition of Speech Decoders and Their Implementation Based on Delayed Evaluation," Proc. APSIPA 2012 (accepted)
(Please cite this paper if you publish a paper using the software)

Discussions for development

A valiation of husky2 by George Bronnikov, Yandex LLC, Moscow. (2012/11/24)
# Thank you for providing the code!

husky-2.1-variation1.tgz source code and sample script
Readme (included in husky-2.1-variation1.tgz)

[back to home]