training end to end using CTC on custom data

186 views
Skip to first unread message

saurabh vyas

unread,
May 12, 2018, 12:33:28 PM5/12/18
to kaldi-help
Hi, I am trying to learn Kaldi, I played with some scripts ( an4 ) and ran the commands given in run.sh, I am interested in end to end systems that make use of CTC Loss, or using encoder-decoder architecture ( with/without attention mechanism ), are there any documentation regarding these ? I have my own data , that is 16K mono .wav files, and sentence level utterances.

Thanks

Daniel Povey

unread,
May 12, 2018, 1:46:27 PM5/12/18
to kaldi-help
> Hi, I am trying to learn Kaldi, I played with some scripts ( an4 ) and ran
> the commands given in run.sh, I am interested in end to end systems that
> make use of CTC Loss

Kaldi doesn't support CTC, mostly because CTC doesn't work very well.
(After extensive experiments with it, I decided not to support it).

> , or using encoder-decoder architecture ( with/without
> attention mechanism ), are there any documentation regarding these ? I have
> my own data , that is 16K mono .wav files, and sentence level utterances.

Nothing like this is supported.

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+...@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/0abffa92-718d-426e-a1aa-d35992e917c6%40googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

saurabh vyas

unread,
May 12, 2018, 1:54:07 PM5/12/18
to kaldi...@googlegroups.com
I see, thanks for your quick reply, Ill try hybrid training then :)

On Sat, May 12, 2018 at 11:16 PM, Daniel Povey <dpo...@gmail.com> wrote:
> Hi, I am trying to learn Kaldi, I played with some scripts ( an4 ) and ran
> the commands given in run.sh, I am interested in end to end systems that
> make use of CTC Loss

Kaldi doesn't support CTC, mostly because CTC doesn't work very well.
(After extensive experiments with it, I decided not to support it).

> , or using encoder-decoder architecture ( with/without
> attention mechanism ), are there any documentation regarding these ? I have
> my own data , that is 16K mono .wav files, and sentence level utterances.

Nothing like this is supported.

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

> 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/0abffa92-718d-426e-a1aa-d35992e917c6%40googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

--
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.

Daniel Povey

unread,
May 12, 2018, 1:57:41 PM5/12/18
to kaldi-help
Hybrid training isn't normally what we recommend. We normally
recommend what the scripts call 'chain' training, which is
lattice-free MMI. It's structurally a little bit similar to CTC,
except that the normalization is per utterance, not per HMM state
(i.e. if CTC is a directed graphical model, LF-MMI is like the
undirected form of that model)... also there are other differences,
such as it's trained against a language model.


http://www.danielpovey.com/files/2016_interspeech_mmi.pdf


Dan

On Sat, May 12, 2018 at 1:54 PM, saurabh vyas <saurabh...@gmail.com> wrote:
> I see, thanks for your quick reply, Ill try hybrid training then :)
>
> On Sat, May 12, 2018 at 11:16 PM, Daniel Povey <dpo...@gmail.com> wrote:
>>
>> > Hi, I am trying to learn Kaldi, I played with some scripts ( an4 ) and
>> > ran
>> > the commands given in run.sh, I am interested in end to end systems that
>> > make use of CTC Loss
>>
>> Kaldi doesn't support CTC, mostly because CTC doesn't work very well.
>> (After extensive experiments with it, I decided not to support it).
>>
>> > , or using encoder-decoder architecture ( with/without
>> > attention mechanism ), are there any documentation regarding these ? I
>> > have
>> > my own data , that is 16K mono .wav files, and sentence level
>> > utterances.
>>
>> Nothing like this is supported.
>>
>> 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+...@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/0abffa92-718d-426e-a1aa-d35992e917c6%40googlegroups.com.
>> > For more options, visit https://groups.google.com/d/optout.
>>
>> --
>> 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/CAEWAuyR_xN8sryw0hMxM-YWUH2x30kT6pWBO6vE8ma%2BvthVBGA%40mail.gmail.com.
>> For more options, visit https://groups.google.com/d/optout.
>
>
> --
> 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/CAHqPSsOmqsOBa_8sDRByW3PpKSQiD8db%2BHQOchQ3nUNpU-Tt5w%40mail.gmail.com.

saurabh vyas

unread,
May 12, 2018, 11:57:03 PM5/12/18
to kaldi...@googlegroups.com
Thanks for the explanation , I'll read the paper for better understanding.

Reply all
Reply to author
Forward
0 new messages