search

kaldi-asr / kaldi

kaldi-asr/kaldi is the official location of the Kaldi project.
http://kaldi-asr.org
Other
13.98k stars 5.3k forks source link

Low latency acoustic modeling using temporal convolution and LSTMs example #4043

Closed witeksosnowski closed 4 years ago

witeksosnowski commented 4 years ago

Hi,

I am trying to build a real time asr and just came across this publication: https://danielpovey.com/files/2017_spl_tdnnlstm.pdf Is there any existing example in kaldi that makes use of this architecture? I was looking at the egs/ami and egs/swbd but could not find any simmiliar example.

Thanks in advance

jtrmal commented 4 years ago

temporal convolution is tdnn, if that's what are you asking

witeksosnowski commented 4 years ago

I meant architecture based on tdnn and lstm that has low latency. It is described in publication I gave. Can you please take a look at it?

jtrmal commented 4 years ago

I'm sorry I don't have time to look for it and it might be well possible the scripts didn't make it to trunk at all (usually we optimize only WER). If you are interested in minimizing latency, you could try what was proposed here: https://groups.google.com/forum/?utm_medium=email&utm_source=footer#!msg/kaldi-help/cjJ8-RWWcao/vZ4i2R0jDAAJ

y.

On Wed, Apr 15, 2020 at 10:13 AM witeksosnowski notifications@github.com wrote:

I meant architecture based on tdnn and lstm that has low latency. It is described in publication I gave. Can you please take a look at it?

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub https://github.com/kaldi-asr/kaldi/issues/4043#issuecomment-614064943, or unsubscribe https://github.com/notifications/unsubscribe-auth/ACUKYX2OGX32E74F2XNX2PTRMW6IDANCNFSM4MIKYQMA .

jtrmal commented 4 years ago

And if you have more questions or follow up questions on this, please use the kaldi-help mailing list y.

On Wed, Apr 15, 2020 at 10:51 AM Jan Trmal jtrmal@gmail.com wrote:

I'm sorry I don't have time to look for it and it might be well possible the scripts didn't make it to trunk at all (usually we optimize only WER). If you are interested in minimizing latency, you could try what was proposed here:

https://groups.google.com/forum/?utm_medium=email&utm_source=footer#!msg/kaldi-help/cjJ8-RWWcao/vZ4i2R0jDAAJ

y.

On Wed, Apr 15, 2020 at 10:13 AM witeksosnowski notifications@github.com wrote:

I meant architecture based on tdnn and lstm that has low latency. It is described in publication I gave. Can you please take a look at it?

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub https://github.com/kaldi-asr/kaldi/issues/4043#issuecomment-614064943, or unsubscribe https://github.com/notifications/unsubscribe-auth/ACUKYX2OGX32E74F2XNX2PTRMW6IDANCNFSM4MIKYQMA .