Lattice is constructed differently in different decoders. Sphinx4 uses straightforward exact lattice construction from a token tree. Sphinx3 and pocketsphinx lattice connections are based on word end times. Which decoder are you interested in?
The good reference for Sphinx3 is a dissertation of Ravi Mosur "Efficient Algorithms For Speech Recognition".
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
It's hard to be sure about that, it's worth to ask on HTK forums.
HTK also has several decoders: HVite, HDecode and HDecode.mod. HVite is slow and HDecode is fast for trigrams. I think HVIte creates exact lattices like Sphinx4 and HDecode creates approximate lattices joined by word end times as described in
The development of the 1994 HTK large vocabulary speech recognition system by Woodland and others
Hi,
Which is the first paper/other reference that described lattice creation in a decoder?
Lattice is constructed differently in different decoders. Sphinx4 uses straightforward exact lattice construction from a token tree. Sphinx3 and pocketsphinx lattice connections are based on word end times. Which decoder are you interested in?
The good reference for Sphinx3 is a dissertation of Ravi Mosur "Efficient Algorithms For Speech Recognition".
Thank you.
I'm interested in Sphinx4 type lattice. Is the same kind of lattice construction method used in HTK also?
It's hard to be sure about that, it's worth to ask on HTK forums.
HTK also has several decoders: HVite, HDecode and HDecode.mod. HVite is slow and HDecode is fast for trigrams. I think HVIte creates exact lattices like Sphinx4 and HDecode creates approximate lattices joined by word end times as described in
The development of the 1994 HTK large vocabulary speech recognition system by Woodland and others
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.28.6618&rep=rep1&type=pdf
Thanks for the pointer.