utilize NNET3 forced alignment in TDNN training

[Lucas Jo]

Hi,

Has anyone have experience to utilize NNET3 forced alignment in TDNN training? 

Can I know the performance difference  b/w using GMM alignment and NNET3 alignment? 

[Daniel Povey]

I don't think it really makes a difference whether you use GMM or nnet3 alignments.

--
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+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/e35ae862-c499-4130-8361-a115bd3846eb%40googlegroups.com.

[Lucas Jo]

Hmm it seem a good news to me. Then I’ll try to use nnet3 alignment for training with additional data

[老师傅]

Actually, I tried your idea, and there is no improvement almost as like Dan said. And if you have a large dataset, nnet3 training is slower than chain, and the alignment is slow too, so maybe you will spend much time on it.

在 2019年8月2日星期五 UTC+8上午9:54:22，Lucas Jo写道：

[Hap Zhang]

Hi, have you tried alignment by chain model, and any  difference?

在 2019年8月2日星期五 UTC+8下午2:27:05，老师傅写道：

[老师傅]

No, I didn't try it.

在 2019年8月9日星期五 UTC+8下午4:12:13，Hap Zhang写道：

[Lucas Jo]

I asked this because I'd like to remove the dependency of GMM model when I try transfer learning outside.  

In my opinion, usually the data for transfer learning is relatively small to the one for full training.

So,  I think it can be useful in some use-case even though it's slow but

Lucas

2019년 8월 9일 금요일 오후 5시 12분 13초 UTC+9, Hap Zhang 님의 말: