Doing forced alignment with custom data

738 views
Skip to first unread message

saurabh vyas

unread,
Sep 25, 2018, 9:10:14 AM9/25/18
to kaldi-help
I know there are a lot of amazing open source force alignment tools, like Gentle, Montreal forced aligner etc., but most of them work for english data, and are trained on english models, I have 2 questions :

1. How does force alignment work in context of kaldi ( word level )? 
 For example, I have an hour of audio, and corresponding text, and I would like to get sentence level alignments, where each sentence consists of words.

2. How can I train my own model in kaldi, for a new language with less data , for forced alignment?
  Also for this, how much data will I need ? 


I have limited Hindi data, and I need to build a custom forced aligner to have a basic system,I can later improve it, once I have more data, Is Kaldi suitable for this ?
Thanks 

Daniel Povey

unread,
Sep 25, 2018, 11:56:35 AM9/25/18
to kaldi-help
You can definitely do this; you would first need to build a system for that language, maybe on a small amount of data from that language (that training data should have short-ish sentences, preferably no longer than 10 seconds; I recommend at least 20 minutes of speech, preferably more).
The script
steps/cleanup/segment_long_utterances.sh  
can then do it.
But I don't claim it will be easy for you to use.  These tools were designed for ASR professionals to use.


--
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/36132590-20e9-44bd-95aa-065857b0e74f%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

saurabh vyas

unread,
Sep 25, 2018, 12:16:27 PM9/25/18
to kaldi...@googlegroups.com
Thanks for your reply, I will have a look at the above mentioned script and try bootstrapping a system with small amount of sentence level data.



Reply all
Reply to author
Forward
0 new messages