Noise Robustness Issue in KALDI

[Hadi Asheri]

Dear Dan,
I am building a GMM-based ASR system with KALDI.
Both training and test utterances are infected by different environmental noises.
You have mentioned in one of your replies to a guy that you have not implemented noise-robust techniques in KALDI because they don't work well with state-of-the-art models.
So, I'd grateful if you could mention some general guides on how handle noise issue with models available in KALDI?
Regards

[Daniel Povey]

Normally fMLLR adaptation works pretty well; this is very hard to beat
by "robust ASR" methods (such papers rarely have fMLLR as a baseline).

Sometimes we artificially add noises to the training data (esp. for
DNN training).

Dan

> --
> 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.
> For more options, visit https://groups.google.com/d/optout.

[vinc...@yahoo.com]

I read somewhere that Sphinx was using some so-called "noise-robust" features with spectral subtraction and temporal masking.

is this what you mention above as not really comparing with fMLLR ?

just curious.

[Daniel Povey]

yes- probably they have not implemented fMLLR. That kind of stuff
helps a lot if you are not doing fMLLR adaptation.