GPU use for decoding

[nikvt...@gmail.com]

Hi,

I was wondering if you plan on using the GPU for the decoding part in the future?

I was looking / trying different toolkits and from what I found, the only projects that are using the GPU for decoding are closed commercials projects..

Thanks,

Nikolay.

[Daniel Povey]

We already support using the GPUs for the neural net computation in
decoding-- it's activated by a flag to programs like
nnet*-latgen-faster, but we don't normally set that to true because
it's not a good use of GPUs (since the graph search also takes time,
and you'd get slowed down by waiting on that).

We don't have plans to work on moving the decoding-graph search to
GPU, because a hassle and because it's not necessary-- with chain
models, we can easily decode faster than real time on the CPU.

Dan

> --
> 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.

[nikvt...@gmail.com]

Thanks for the super fast reply!

I'm interested in using the gpu for multiple (as much as possible) concurrent real-time asr tasks, that's why I was asking.

Nikolay.

понедельник, 9 мая 2016 г., 12:17:32 UTC-7 пользователь Dan Povey написал:

[qiang gao]

hi,

I am interested in your topic, and do you have any progress.

在 2016年5月10日星期二 UTC+8上午3:48:00，nikvt...@gmail.com写道：

[Daniel Povey]

No, there are still no plans to use the GPU for graph search in Kaldi.