How to use the pre-trained model

[Tien Hong Lo]

Hi Everyone,

I want to use the pre-trained model on small corpus.(decode directly and retrain)

And I noticed that ASpIRE Chain model form http://kaldi-asr.org/models.html .and download it.

But I don't know how to use it ...

Should I use the same ivector extractor, mfcc_hires_conf ?

[Daniel Povey]

Yes, you should use the same ivector extractor and MFCC config,
otherwise you'd get nonsense results.

> --
> 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 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/61936245-273f-4f1e-91d7-fb3b26e2bae1%40googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[Tien Hong Lo]

Thanks for your quickly reply!

However, I have two questions ...

1. In archive, I'm not found that "final.ie.id" of aspire/exp/chain/tdnn_7b and "conf" of aspire.

Where should I find it?

2. If I want to use the model (tdnn_7b/final.mdl) as initial model and retrain it on our small corpus.

For "ivector extractor" and "MFCC config", I should use our corpus setting or follow by aspire?

In your experience, which is more theoretically correct?

Dan Povey於 2018年5月27日星期日 UTC+8上午12時57分05秒寫道：

[Daniel Povey]

>
> However, I have two questions ...
> 1. In archive, I'm not found that "final.ie.id" of aspire/exp/chain/tdnn_7b
> and "conf" of aspire.
> Where should I find it?

Ignore the warnings about that. It was not present in older models.

> 2. If I want to use the model (tdnn_7b/final.mdl) as initial model and
> retrain it on our small corpus.
> For "ivector extractor" and "MFCC config", I should use our corpus setting
> or follow by aspire?
> In your experience, which is more theoretically correct?

Unless you leave them the same, it won't work at all.


Dan

> https://groups.google.com/d/msgid/kaldi-help/340db7a8-0fb5-4ca2-835d-90e5892663e9%40googlegroups.com.

[Tien Hong Lo]

Thanks for your reply!

I want to use "MFCC config" the same as aspire, but I'm not found it in archive.

Is this "MFCC config" the same as the example provided by github (kaldi/egs/aspire/s5/conf)?

If so, I would like to ask questions about the different sampling frequencies. 

My corpus is 16kHz, by contrast,  aspire is 8kHz. 

Is there any way to use it? or should I consider other models?

Dan Povey於 2018年5月27日星期日 UTC+8上午1時55分01秒寫道：

[Daniel Povey]

Yes, it will be the same as what's in the repository.
If you have 16kHz data you have the choice to use an 8kHz model like
aspire (you can set --allow-downsample=true in mfcc.config to make it
do the downsampling automatically); or you can use a model built for
16kHz, such as one from Librispeech.

Dan

> https://groups.google.com/d/msgid/kaldi-help/fde99bb2-b0b4-4611-a41a-d38e73de0e4d%40googlegroups.com.

[Tien Hong Lo]

Thanks, 

I will try that ! 

Dan Povey於 2018年5月28日星期一 UTC+8上午9時06分40秒寫道：