Kaldi MFCC

[jamarshon]

We want to write compute-mfcc-feats for compliance.kaldi.mfcc in torchaudio

[jamarshon]

current result on 70 files. those should be max error not mse

max relative error for fbank was 0.081825569 so seems suspicious here

[jamarshon]

Found a bug. Now the code has the following statistic for the 70 files:

Max absolute error: 0.00029754638671875
Max relative error: 5.5

The 5.5 max relative error has a max absolute error of 5.245208740234375e-06. For file

mfcc-2.0603-1.7529-0.6875-0.6250-7900-false-5086-4-0.74-false-false-true-false-true-false-4-59.0671-7431-7014-1.0000-hamming.ark

which is acceptable. Now the graph looks more like expected 😄

[vincentqb]

I'm glad to see that there are over 2^6 test files to test all true/false combinations :)

Thanks for providing error graphs. This looks good to me!

[vincentqb]

I see a commit about the README, but I don't see changes in README files. Did you mean to add one?

[jamarshon]

I added a small change to the README to say mfcc along with the related kaldi features we currently have
https://github.com/pytorch/audio/pull/228/files#diff-04c6e90faac2675aa89e2176d2eec7d8R25

[vincentqb]

I added a small change to the README to say mfcc along with the related kaldi features we currently have
https://github.com/pytorch/audio/pull/228/files#diff-04c6e90faac2675aa89e2176d2eec7d8R25

Right, I see what that meant now. Thanks!

[vincentqb]

I'm glad to see that there are over 2^6 test files to test all true/false combinations :)

As discussed with @cpuhrsch:

• They don't cover all cases since they are picked randomly.
• We also should get in the habit of building more thorough edge cases (e.g. empty tensors? all zeros?).
• Long list of parameters can be passed easily using locals.

[jamarshon]

new statistic over 145 files (same overall shape just more points on the left region) + added test for empty tensors