how to compute cmvn in online decoding

[Yu Beomgon]

How can I use cmvn in decoding?

cmvn is calculated with trained data(using spk2utt etc)

but in test, only audio file we have.

in the guide, below explanation I found, but I dont understand and how to use.

The basic solution we use is to do "moving-window" cepstral mean normalization. We accumulate the mean over a moving window of, by default, 6 seconds (see the "--cmn-window" option to programs in online2bin/, which defaults to 600). The options class for this computation, OnlineCmvnOptions, also has extra configuration variables, speaker-frames (default: 600), and global-frames (default: 200). These specify how we make use of prior information from the same speaker, or a global average of the cepstra, to improve the estimate for the first few seconds of each utterance. The program apply-cmvn-online can apply this normalization as part of a training pipeline so that we can can train on matched features.

https://kaldi-asr.org/doc/online_decoding.html

please tell me how to do and if there is an example, give me the link.

thanks

[Daniel Povey]

If you trained the model with the normal scripts (i.e. not the scripts in steps/online), you should calculate the CMN with the scripts, via `compute-cmvn-stats` and then applying the shift with `apply-cmvn`.  The model likely wouldn't be compatible with moving-window CMN if it was trained with regular CMN.  As I said before, you need to look at the decoding command line and trace back to the log files with the commands where the various things it uses were created; and utt2spk and spk2utt would normally be dummy files with just a one-to-one map from utterance to/from speaker.

--
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/5b2b7546-24db-45d2-a68e-ae61c2f58418%40googlegroups.com.