ivector

[tumma...@gmail.com]

Hello everyone,

how can I connect the normal features and online_ivector together to train the neural network? i found in tedlium/local/nnets/run_tdnn.sh, the 100 dimensions' ivector with 40 dimensions' features are used to train the "tdnn" model. Could I use "append-vector-to-feats" to connect the 40 dimensions' features and 100 dimensions' ivector? And I found in the "tdnn" script, the ivector is online_ivector, and the online_ivector are extracted from every 10 frames. So the normal features and the online_ivector are not one-to-one corresponded, right? Someone could give me some advice?

Regards.

Y

[Daniel Povey]

If you are trying to reproduce Kaldi results in another toolkit, I recommend to try to reproduce the results without ivector for now.  You can get the results without ivector by removing the ivector-related things from the xconfig and everything after that in the xconfig file.  That includes replacing

 input=Append(-1,0,1,ReplaceIndex(ivector, t, 0))

with

 input=Append(-1,0,1)

You likely won't be able to even get close to the Kaldi-based results, since you don't have experience with this kind of thing, so the 10% that you can get from ivectors is just a rounding error.

Anyway it's not as simple as just appending them to the features; that lda-like transform does help, and it's a little complicated to explain briefly.  At the very least you would need to scale the ivector by 0.1 or so to prevent it overwhelming the time-varying features.

Dan

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
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+unsubscribe@googlegroups.com.
To post to this group, send email to kaldi...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/c802fd7d-1013-4d78-9cde-b56edafdb1d8%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[刘叶]

Thanks, dan. I get your meaning, and I will try without ivectors.

Best regards,

Y

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.

To post to this group, send email to kaldi...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CAEWAuySjNwY_7AcdcQjXE-qi%3D%3D6vZKPVCfpNhJJquxpkvKPpXw%40mail.gmail.com.