Speaker identification UBM training.

[kri...@cogknit.com]

Hi All,

I have a dataset of 30 people speaking to microphone for 3 mins each and i want to build a Speaker identification model ( For now GMM-UBM). I tried sre10 v1 recipe by taking 25 speakers for UBM training and 5 speakers for enrollment. I got 40% EER which i think is very bad. I am thinking it is because of i had very small data for UBM training. I have access to WSJ and BABEL data. Is WSJ is better dataset for UBM training for my problem?  or BABEL is better dataset for UBM?

[Daniel Povey]

You need much more data than that: preferably at least a thousand hours, but hundreds might be OK.

It's generally important to have data that covers your application domain of interest, and contains the same speakers from multiple recording conditions.

I don't think either WSJ or BABEL data would be very good, but they might be better than nothing.

On Fri, Feb 9, 2018 at 12:47 AM, <kri...@cogknit.com> wrote:

Hi All,

I have a dataset of 30 people speaking to microphone for 3 mins each and i want to build a Speaker identification model ( For now GMM-UBM). I tried sre10 v1 recipe by taking 25 speakers for UBM training and 5 speakers for enrollment. I got 40% EER which i think is very bad. I am thinking it is because of i had very small data for UBM training. I have access to WSJ and BABEL data. Is WSJ is better dataset for UBM training for my problem?  or BABEL is better dataset for UBM?

--
visit http://kaldi-asr.org/forums.html to find out how to join.
---
You received this message because you are subscribed to the Google Groups "kaldi-developers" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-developers+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Krishna D N]

It was helpful feedback... Thank you

You received this message because you are subscribed to a topic in the Google Groups "kaldi-developers" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/kaldi-developers/h0Cm5j6VGGA/unsubscribe.
To unsubscribe from this group and all its topics, send an email to kaldi-developers+unsubscribe@googlegroups.com.