- fixed serious bug in the extractor, ReplayGain value was always 0
- fixed crash when using deleteNetwork on some AlgorithmComposites
- added range description for parameters, also does automatic range checking when
configuring an algorithm
- added new python unit tests and streaming C++ unit tests
- did a white-box review of Essentia's most used algorithms, they now:
- have input/output/parameter names which conform better to the guidelines
- do input validation
- some descriptors output by the streaming_extractor have been fixed and now have
different values. Here is the list of changes:
- MFCC are slightly different (negligible), because the bands limits were not
correctly taken before (off by 1 error)
- Barkbands were overlapping before, not anymore so some bands are the same,
some others have smaller energy values.
differences are significative, and are more significative in lower bands
(because we have lower resolution in low frequencies)
Descriptors dependent on barkbands:
- barkbands_kurtosis -> changes a lot, but 4th order moment -> too sensitive to noise
- barkbands_skewness -> changes a lot, but 3th order moment -> too sensitive to noise
- barkbands_spread -> changes by ~1%, quite stable
- spectral_crest -> changes a lot -> uses the max which is way too sensitive to noise
- spectral_flatness_db -> changes a little (~5%)
- average loudness: very small deviation (<1%)
due to the framecutter not returning exactly the same number of frames (off by 1)
this difference does not appear on all the files (some have the exact same value)
- tempotap (and dependencies) has negligible variations (<0.01% for all dependencies)
- beatsloudness_bass varies a lot because we used energybandratio rounds to nearest
bin instead of truncating, hence we're not taking into account the bin nº 0,
which has lots of energy