biglm or lattice+rescore

[4151...@qq.com]

Hi,

     There are two algorithms in Kaldi to perform the decoding with large language models (i.e., 60G ngrams). Which one is the popular trend? can you give us some suggestions. We plan to rewrite the decoder without Kaldi.

[Daniel Povey]

lattice + rescoring.
biglm turned out to be quite slow

> --
> You received this message because you are subscribed to the Google Groups
> "kaldi-help" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to kaldi-help+...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[4151...@qq.com]

I did some experiments, and I found the "TableCompose" process cost much time, e.g. around 2 second to compose a lattice (generated by a 10 second wav) with a 600M LM. It seems to be slow.

在 2016年12月30日星期五 UTC+8上午10:51:32，Dan Povey写道：

[Daniel Povey]

It depends how deep the lattices were (e.g. how big the beam was).
Also make sure you don't include initialization/loading time (e.g. the
time to load the FST) in that time.
And 'carpa' rescoring may be faster.

[4151...@qq.com]

Thanks, I did not include the loading time (it usually costs more than 10 seconds). The lattice generation uses beam = 15, max-active=7000, and the wav length is around 10 seconds.

在 2017年1月4日星期三 UTC+8下午3:37:42，Dan Povey写道：

[Daniel Povey]

There is time taken in building certain internal tables, that will
only get taken for the first utterance, so you won't see the best
speed if you run it on just one utterance.