robust feature

[har...@gmail.com]

Hi Dan

 

Do you have experience about robust feature ?

Now kaldi use MFCC or PLP feature

but some papers mention the accuracy of using mfcc is low in the low SNR environment

 

My question is

Which feature is robust and computation complexity is reasonable for real environment use ?

kaldi have any plan to add robust feature ?

like ETSI-AFE or PNCC .....

 

Thanks

 

[Jan Trmal]

If I recall correctly, we tried PNCC for babel data (which is pretty noisy) and we didn't see any improvement.

y.

--
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.

[Niko Moritz]

From my experience, the difference between PLP and MFCC is negligible, so I prefer the more easy and flexible implementation, which is MFCC.
PNCCs are strong if you train with clean data and test with noisy data but their advantage seems to dissolve in multi condition training.
The ETSI AFE works well on Aurora data sets, since their speech enhancement block is tuned to these noise types but in more realistic noise conditions they also lose strength.
fbank features with splicing is a good solution, because you put more freedom to the DNN and with enough training data it can extract the essential information by itself by learning good modulation filters. If you like to put more knowledge of modulation filtering into the feature extraction, Gabor, MRASTA, TRAP-DCT and AMFB features seem to do a good job.

[har...@gmail.com]

Thanks for everyone share experience

 

How about multi-condition training ?

only use mfcc without de-noise algorithm.

 

har...@gmail.com於 2015年9月14日星期一 UTC+8下午3時33分14秒寫道：