Add a real example test to PitchYin and PitchYinFFT unittests

[dbogdanov]

Current test only checks if the algorithms' output has changed. A Mozart piece is used, but the reference pitch values are the values PitchYin / PitchYinFFT algorithms have output for it in the past.

This test now fails after updates introduced by #376 and should be updated with new "fake" reference values.

Add a new test with real reference values. Several monophonic recordings of various instruments playing scales from GoodSounds might work.

Ideally, we need three types of tests:

• regression test on real audio to detect if algorithm output has changed (before vs after)
  • compare pitch estimation with previously registered estimation
• test on synthesized pure tones (#230)
• test on synthesized instruments playing a scale
  • similar to pure tones but with multiple instruments synthesized using TiMidity
• test on real instruments playing a scale
  • create reference annotation manually
  • the estimated pitch is allowed to be within some deviation due to the nature of real instruments

It will be very helpful for Q/A purposed if these tests can generate reports with plots.

[oriolromani]

For bullet 4 we can create a reference annotations in a semi manual way as it was done for the pitch annotations of MedleyDB. They use Tony software to first estimate the fundamental frequency using pYin algorithm and then refine it manually. Here is an example plot of a comparison with PithcYinFFT essentia's algorithm.

[dbogdanov]

To summarize our previous discussions:

• Tests will check if the estimated pitch is close to the ground-truth annotation for non-silent segments
• Silent segments between notes should are annotated with 0Hz in the ground truth, and should correspond to low confidence values in corresponding pitch estimation.