Difference between online and offline features

[Weiji Zhuang]

Hi all,

I trained a model with offline features (extracted with compute-mfcc-feats), and the offline decoder tested well (feed feds.scp).

But when I tested with the online decoder (feed wav.scp), The result is worse.

I printed the features and found that the two features are not the same, as follows:

online:hires_mfcc

Extract with online-nnet2-feature-pipeline

*features_[0].data_@40 108.058304 -32.2430229 -6.12382746 -35.8702354 ...

*features_[1].data_@40 127.713455 -46.533802 -7.89504623 -38.0356674 ...

*features_[2].data_@40 127.917244 -47.0403175 -12.6836214 -39.5548363 ...

*features_[3].data_@40 124.277847 -40.7552872 -8.98573685 -47.4460907 ...

*features_[4].data_@40 122.870132 -43.1498299 -20.4687481 -61.6906433 ...

*features_[5].data_@40 122.521553 -44.9273338 -26.5097141 -59.5551262 ...

*features_[6].data_@40 121.453888 -45.7891426 -25.7881966 -58.3754578 ...

*features_[7].data_@40 120.90654 -47.3259888 -28.4465427 -53.8251305 ...

*features_[8].data_@40 121.196655 -43.5808029 -26.9738884 -57.0203209 ...

*features_[9].data_@40 118.689552 -39.5210533 -29.2908669 -59.9201622 ...

*features_[10].data_@40 114.087212 -40.0591164 -35.7206039 -59.328743 ...

*features_[11].data_@40 109.121292 -32.3833084 -30.9196854 -49.0763321 ...

*features_[12].data_@40 97.9033203 -30.993185 -23.445219 -29.7657738 ...

*features_[13].data_@40 87.8285675 -38.1543999 -18.3624916 -21.0542469 ...

*features_[14].data_@40 85.3344116 -37.3253937 -12.7687273 -16.2427673 ...

*features_[15].data_@40 83.5251617 -36.7737503 -11.4216547 -16.8593807 ...

offline:hires_mfcc

Extract with compute-mfcc-feats

feats.ark:frame0 108.144 -32.21425 -6.269695 -36.08073 ...

feats.ark:frame1 127.5763 -46.86416 -7.838353 -38.27585 ...

feats.ark:frame2 127.9172 -46.86416 -12.54433 -39.73927 ...

feats.ark:frame3 124.1671 -40.58563 -9.014847 -47.78802 ...

feats.ark:frame4 122.8035 -43.37609 -20.38762 -61.69043 ...

feats.ark:frame5 122.4626 -44.77132 -26.65874 -59.49531 ...

feats.ark:frame6 121.4398 -45.46893 -25.87519 -58.0319 ...

feats.ark:frame7 120.758 -47.56178 -28.61761 -53.64167 ...

feats.ark:frame8 121.0989 -43.37609 -27.05051 -57.3002 ...

feats.ark:frame9 118.7125 -39.1904 -29.40116 -60.22702 ...

feats.ark:frame10 113.9396 -39.88801 -35.66955 -59.49531 ...

feats.ark:frame11 109.1668 -32.21425 -30.96826 -49.25143 ...

feats.ark:frame12 97.69287 -30.81902 -23.52455 -29.49538 ...

feats.ark:frame13 88.0256 -38.49278 -18.4268 -20.71492 ...

feats.ark:frame14 85.26353 -37.09755 -12.93649 -16.32468 ...

feats.ark:frame15 83.42215 -37.09755 -11.36783 -17.05639 ...

They use the same mfcc_hires.conf file. Why offline features (extracted with compute-mfcc-feats) and online features (extracted with online-nnet2-feature-pipeline) are inconsistent?

Best Regards,

Weiji Zhuang

[Daniel Povey]

Might be due to compression if you got one of them from disk.

--
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/78fde05f-1c2f-4469-987a-0655634be2b5%40googlegroups.com.

[Weiji Zhuang]

Cool!   after I changed the --compress option to false, the online and offline are consistent.

I want to know, does the problem of inconsistent performance after compression appear in all recipes?

The performance degradation of the mfcc feature is still relatively small. When I changed the feature to a spectrogram, the performance loss was completely unacceptable.

[Daniel Povey]

Spectrograms are harder to compress than MFCCs (for that algorithm) because they are more correlated.

On average I don't think there is performance loss, when using MFCCs.

--

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/1fbbee3b-af74-4ace-82a6-3c4f87694f81%40googlegroups.com.