Migrate Kaldi tests

[vincentqb]

We used to test kaldi against a saved kaldi output here, but we now have an infrastructure to run kaldi in our test environment here.

We would like to migrate the tests from "saved kaldi" to "live kaldi".

• test_spectrogram
• test_fbank
• test_mfcc
• test_mfcc_empty
• test_resample_waveform
• test_resample_waveform_upsample_size
• test_resample_waveform_downsample_size
• test_resample_waveform_identity_size
• test_resample_waveform_downsample_accuracy
• test_resample_waveform_upsample_accuracy
• test_resample_waveform_multi_channel

[vincentqb]

We should also have CI catch errors like #613 on gpu

[bhargavkathivarapu]

@vincentqb , For below kaldi compliance tests, Do we need to write a compatibility test cases for each saved files configurations(function args in filenames) in test/kaldi or one test configuration is enough for each

• test_spectorgram
• test_fbank ( one configuration already there in compatibility tests )
• test_mfcc
• test_resample_waveform

For remaining tests , they are not dependent on the saved kaldi files in test/kaldi ( though they are referencing a external .wav file , they are not comparing its contents ) , we can directly move them to compatibility without much changes

[vincentqb]

The tests should cover the same cases. Is that what you meant?

[bhargavkathivarapu]

The tests should cover the same cases. Is that what you meant?

@vincentqb yes , Like for test_fbank there are 97 kaldi files , each file name is variation of kaldi fbank function arguments

Two example file names out of 97 files
- fbank-0.0939-4.5062-1.0625-0.6875-1841-true-479-5-0.84-true-true-true-true-true-true-true-false-1832-1824-1.0000-hanning.ark

- fbank-0.1660-1.7875-1.1250-0.5000-4999-true-1740-6-0.29-true-false-true-false-false-true-false-true-4587-2289-1.0000-povey.ark

So we need to keep all these 97 configurations( function argument variations) in a text file ,
In the test we read this text file, and for each configuration we generate kaldi output on the fly and compare with torch output
Like this right ? (or instead of all 97 we can test for 1 configuration only ?)

[vincentqb]

So we need to keep all these 97 configurations( function argument variations) in a text file ,
In the test we read this text file, and for each configuration we generate kaldi output on the fly and compare with torch output

Yes, that is what I mean. We need to maintain the same coverage.